Are heterogeinity and conflicting preferences no longer a problem? Personality-based dynamic clustering for group recommender systems

Are heterogeinity and conflicting preferences no longer a problem? Personality-based dynamic clustering for group recommender systems Patrícia Alves a , b , c , * , Andr´ e Martins b , Francisco Negr˜ ao b , Paulo Novais a , c , Ana Almeida b , c , Goreti Marreiros b , c a ALGORITMI Research Centre, School of Engineering, University of Minho, Campus de Gualtar, 4710-057 Braga, Portugal b GECAD, Superior Institute of Engineering, Polytechnic of Porto, Rua Dr. Ant´ onio Bernardino de Almeida, 431, 4249-015 Porto, Portugal c LASI, School of Engineering, University of Minho, Campus Azur´ em, 4800-058 Guimar˜ aes, Portugal ARTICLE INFO Keywords: Group Recommender System Personality Dynamic Clustering Multi-Agent Microservice Cold-Start Leisure Tourism ABSTRACT The complexity associated with groups of tourists led to the emergence of Group Recommender Systems (GRS) for tourism. But if generating recommendations for small groups is a complex task, to provide them to large and occasional groups is even more. This complexity is especially due to the group’s heterogeinity, conflicting preferences, the information overload found on the internet and the tourists’different ways of coping with the information, hindering the recommendation process from the users’profile construction to the final recommendation of a list of points of interest to visit. In this work, we show how we tackled the identified issues in a GRS prototype, Grouplanner, including the cold-start problem, by predicting the tourists’preferences based only on their personality and dividing the main group into subgroups of similar personality; by using a Multi-Agent Microservice; a novel dynamic clustering algorithm, d-means, adapted from the k-means algorithm, that does not need to know the number of clusters a priori; and association rules. Using a personality dataset of n=100k users, the proposed d-means algorithm was tested against two baselines (k-means and k-means++), showing better results in the clustering quality and scalability. We were also able to determine a large set of association rules to refine the recommendations, although further improvements are needed. To test the Grouplanner prototype, a simulation with real users (n=35) was conducted. The results showed the subgroups formed were very compact, revealing a very good clustering quality, with an average silhouette of s=0.91. 11 of the 15 proposed tourist preferences were successfully predicted and used for the preliminary recommendation lists, being 92 % of the participants satisfied with the individual recommendations and 96 % with the group recommendations. 1. Introduction Leisure tourism is a very complex area (Alves, Carneiro, Marreiros, & Novais, 2019; Deli´ c, Nguyen, &Tkalˇ ciˇ c, 2020; Garcia, Sebastia, & Onaindia, 2011; Neidhardt, Schuster, Seyfang, &Werthner, 2014; Sertkan, Neidhardt, &Werthner, 2019) as it involves sensorial experiences and emotions, being subject to many variables. This complexity aggravates when groups of tourists, like excursions of casual groups of tourists, are involved, increasing the heterogeinity and conflicting preferences (Boratto &Carta, 2010), which is not made easier with the excessive information found on the internet. Where to go? Which tourist attractions are more compelling for a certain group of tourists? Group Recommender Systems (GRS) for tourism came to tackle that complexity and information overload, helping in the choice of a destination, accommodations, points of interest (POI) to visit, etc. However, the GRS for tourism found in literature are limited, either having timeconsuming and intrusive configurations, a great amount of data to build the tourists’initial profile, complex preference aggregation strategies, or are not concerned about how the tourists feel during the excursion and in the group, like reducing the sense of misplacement, increase the group members socialization and the sense of companionship and belonging to the group, or how their limitations like fears/ phobias or disabilities impact the group satisfaction. How can these limitations be overcome? * Corresponding author at: GECAD, Superior Institute of Engineering, Polytechnic of Porto, Rua Dr. Ant´ onio Bernardino de Almeida, 431, 4249-015 Porto, Portugal. E-mail addresses: [email protected] (P. Alves), [email protected] (A. Martins), [email protected] (F. Negr˜ ao), [email protected] (P. Novais), amn@isep. ipp.pt (A. Almeida), [email protected] (G. Marreiros). Contents lists available at ScienceDirect Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa https://doi.org/10.1016/j.eswa.2024.124812 Received 4 January 2024; Received in revised form 15 June 2024; Accepted 12 July 2024 Expert Systems With Applications 255 (2024) 124812 Available online 19 July 2024 0957-4174/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/bync-nd/4.0/ ). As first identified by Jameson (2004), and as well known in literature, the four main challenges of GRS are: (i) how to obtain the group members’ individual preferences and if they should be known and analyzed by the other members; (ii) how the users’preferences should be combined, in a nonmanipulable manner, to generate recommendations that satisfy all the group members; (iii) how to effectively present suitable and explainable recommendations for each group member; (iv) how to support the group members in deciding and negotiating the final recommendation. As evidenced in our previous work (Alves, Martins, Saraiva, et al., 2023; Alves et al., 2020) and in the existing literature, personality has a strong influence on the users’preferences in a variety of domains (Cantador &Fern´ andez-Tobías, 2014; Ferwerda, Schedl, &Tkalcic, 2015; Martijn, Conati, &Verbert, 2022; Rawlings &Ciancarelli, 1997; Yee, Ducheneaut, Nelson, &Likarish, 2011) and is demonstrating to provide better results than standard approaches (Nunes, Cerri, &Blanc, 2008; Tondello, Orji, &Nacke, 2017). This can be the leverage needed to improve group recommendations, as the correlation between the users’ personality and preferences can match up users with similar interests helping to minimize the group’s heterogeinity and conflicting preferences, mitigating the cold-start problem. As human personality is considered stable over time (McCrae &Costa Jr, 1997), the users’personality only needs to be assessed once, being an advantage that can be seized and reused across different domains. This work aims to help solve the identified limitations by automatically providing recommendations of tourist attractions/activities to visit/perform at a certain destination, along with a justification, while caring for the tourists’needs, without the need for previous data to build the tourists’initial profile and the first recommendations, solving at the same time the cold-start problem, the groups’heterogeinity and conflicting preferences. That is, the aim is to provide initial recommendations that satisfy the majority of group members with the minimum necessary information, just by knowing the tourists’personality, and avoid the emotional contagion phenomenon usually associated to groups (Delic &Masthoff, 2018; Deli´ c et al., 2020), being, to the best of our knowledge, the firsts to do so (Sharma, Banwala, Elouaghzani, & Katarya, 2023). This paper unfolds how we implemented the personality vs tourist preferences models in a mobile GRS prototype, Grouplanner, and modified the algorithms proposed in our previous work (Alves, Martins, Saraiva, et al., 2023), to solve the identified issues. This new approach dynamically creates (in real time) personality-based clusters when the tourists register in the app, eliminating the need for users to rate POI to create the initial clusters as previously proposed (Alves, Martins, Saraiva, et al., 2023), leading to an alternative algorithm and solving the cold-start problem. This paper materializes a comprehensive study of the work done so far, so it can be understood as a whole, being the prototype functioning briefly demonstrated inAlves, Martins, Novais, and Marreiros (2023). 1 To validate the proposed d-means algorithm it was compared with two baselines: k-means and k-means++. To validate the prototype, it was tested in a simulation with real users and the results analyzed. In short, this paper has the following new contributions: •A personality-based dynamic clustering algorithm, d-means, is proposed, being a novel strategy for aggregating preferences, that aggregates tourists with high similar personalities (≥0.80) using an adaptation of the k-means algorithm, automatically creating the best clusters in real-time as the users register in the application, being to the best of our knowledge, a novelty in GRS; •As the clustering is based on the tourists’personality high similarity, we eliminated the need to know the number of initial clusters as well as the need for previous interaction data and ratings from users. This strategy solves the cold-start problem, as the clusters’members are automatically assigned with a predicted preliminary preference score in categories of tourism and travel-related preferences &concerns immediately after registering, allowing a preliminary recommendation list. The clusters are used to form subgroups of tourists with similar interests when requesting a recommendation for large/ occasional groups of tourists, as we believe it will be more beneficial to them than to stay in the main (heterogeneous) group; •Use of the Apriori algorithm to find rules in the existing clusters, according to the tourists’profile and travel history. The rules are used to create lists of POI to include or exclude in the recommendation lists. This information is very useful for the recommendation engine at the time of assigning the points of interest, so that and other similar POI are included or not in the recommendation list for a subgroup, ameliorating the initial recommendations; •Consideration of weather conditions using data retrieved from an external weather API. This means when a recommendation is requested, the recommendation engine will check for suitable POI according to the weather conditions on the excursion dates, updating the recommended list from 5 days before the excursion begins until it ends. Beside the referred contributions, we are also working on a proof of concept to gamify the GRS prototype to remove time-consuming and intrusive configurations, such as the personality questionnaire used to assess the users’personality, by using short duration minigames to implicitly acquire the tourists’personality (Alves et al., 2024). With this, we intend to show that the users’personality traits can be acquired unobtrusively just by using minigames and without the need for large amounts of data, and therefore machine-learning algorithms, to gather it. The remainder of the paper is structured as follows: Section 2 describes the background and some related work on GRS for Tourism, Agent-based RS, and clustering techniques used in GRS, along with some reflections; Section 3 presents the methodology used; Section 4 describes the conducted experiments to validate the proposed d-means algorithm and the performed simulation to test the GRS prototype. Section 5 shows the obtained results and respective analysis; and finally, Section 6 reflects on the contents addressed in the paper and describes what will be done as future work. 2. Background and related work Although many RS for tourism exist in literature, implementing and comparing different types of Machine-Learning and Decision-Making algorithms for different stages of the recommendation process, ranging from a single type of tourist activity to a broader range, not many are for groups (Amato, 2023; Blanco, Ge, &Pitner, 2023; Djebali, Gabot, &Guerard, 2023; El Majjodi, Elahi, El Ioini, &Trattner, 2020; Gavalas, Konstantopoulos, Mastakas, &Pantziou, 2014; Kuo &Li, 2023; Nadeem et al., 2022; Vathis, Konstantopoulos, Pantziou, &Gavalas, 2023; Yalcin, Ismailoglu, &Bilge, 2021). In the Section 2.1, we present some relevant works. 2.1. Group recommender systems for tourism Some of the first GRS used simpler techniques at the time, like INTRIGUE (Ardissono, Goy, Petrone, Segnan, &Torasso, 2003), which was one of the first GRS, that helps find sightseeing destinations and itineraries in Italy based on manual configurations made by a responsible tourist to characterize the members representing the excursion group. The configurations are then used to divide the group into subgroups of similar preferences, being the recommendations given to each subgroup along with some explanations. Some authors use preference aggregation and incompatible preferences management like Travel Decision Forum (Jameson, 2004), e-Tourism (Garcia, Sebastia, Onaindia, &Guzman, 2009), and MyFoodGRS (Emamgholizadeh, Deli´ c,&Ricci, 2024), or average ratings and/or critiques like CATS (McCarthy, McGinty, Smyth, & 1 More details can also be found at https://www.gecad.isep.ipp.pt/groupla nner/dissemination.html. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 2 Salam´ o, 2006; McCarthy, Salam´ o, et al., 2006), Where2eat (Guzzi, Ricci, &Burke, 2011) or the chat-based mobile GRS STSGroup (Nguyen &Ricci, 2018). Abolghasemi, Engelstad, Herrera-Viedma, and Yazidi (2022) proposed a GRS based on pairwise preferences and on the Thomas-Kilmann conflict resolution styles (Thomas, 2008) that can be used in large groups and in different domains. The TKI styles and pairwise preferences are used to simulate the negotiation between the group members, where the members influence each other to reach a consensus on the ratings according to their weights on the TKI styles. The GRS proposed by Park, Park, and Cho (2008) uses a Bayesian network to model the user’s individual preferences based on the user’s context and profile. The Analytic Hierarchy Process from multi-criteria decision-making is then used to provide the most suitable recommendations of restaurants considering the modeled individual preferences. Hootle+(´ Alvarez M´ arquez &Ziegler, 2016) is a hotel GRS prototype where each group member chooses the features they want the hotels to have, and how much important a feature is. The chosen features are then presented to the group, which decides to accept them as a filter or not, creating the group’s preference model. Using content-based filtering, the preference model is compared to the hotel’s attributes in the database, recommending the hotels with the highest scores. The final recommendation is reached upon the group members negotiation and voting. A similar approach was made by Marques, Respício, and Afonso (2016) for a mobile GRS for choosing a restaurant. New approaches are being considered, like using location-based social networks (LBSN), like Sojahrood, Taleai, and Cheng (2023) who developed a POI GRS based on LBSN, with a dataset retrieved from the Foursquare Swarm, that uses a switching hybrid group recommendation algorithm for different types of groups: for ephemeral groups, they use a group-based approach; for persistent and homogeneous groups, they use a simple aggregation-based method; and for persistent and heterogeneous groups, they execute the improved-based method; outperforming the baselines in the three types of groups. Other examples of GRS that use LBSN to recommend the most suitable POI are the works of AyalaG´ omez, Dar´ oczy, Mathioudakis, Benczúr, and Gionis (2017) and Khazaei and Alimohammadi (2019). A very recent approach is the one of Delic, Emamgholizadeh, Nguyen, and Ricci (2024), who use a chatbot to support the group discussions, preference elicitations, and item rankings, to provide a fair recommendation using the average without misery preference aggregation strategy. As can be seen from the existing literature, the GRS authors are attempting to find better strategies to provide more accurate and satisfying group recommendations, trying to involve the user in the recommendation process, and find better preference aggregation and negotiation strategies. However, the research in the area is slowly evolving, still having important open issues (´ Alvarez M´ arquez &Ziegler, 2016; Alves et al., 2019; Deli´ c et al., 2020; Masthoff, 2015; Ricci, 2022), such as the best approach (a) to know the users’preferences (implicitly or explicitly), (b) to provide the first recommendation without the need for the users interaction, negotiation or feedback (lack of ground-truth data), (c) to combine the individual preferences to provide more accurate and satisfactory group recommendations, or if instead of aggregating preferences, the group members interactions should be mediated, (d) to satisfy the needs of the group according to the context, (e) to increase the sense of companionship and belonging to the group, (f) diminish the sense of fear/being lost or (g) to support the group members in reaching a faster and more satisfactory final choice without the social contagion phenomenon. One of the RS main purposes is to provide automatic recommendations, but, to the best of our knowledge, none, except one (Park et al., 2008), provides a preliminary automatic recommendation for the group, depending exclusively on the users interaction and feedback on the existing features or POI to provide the first recommendations. Also, according to Sertkan et al. (2019), few studies focus on the tourists’ personality. These are two of the open issues, (a) and (b), we propose to solve, by automatically providing initial recommendations without the users explicit desired preferences or feedback on the existing POI, using only the tourists’personality, which also solves the cold-start problem. The works of Sertkan, Neidhardt, and Werthner (2020) and Neidhardt, Seyfang, Schuster, and Werthner (2015) are examples of an interesting way of implicitly acquiring the tourists’profile, the former by uploading up to 7 pictures the user considers to represent his preferences which can then be re-ranked to build the touristic profile, that can also be adjusted, to obtain destination recommendations. This strategy implies some work from the user to find the best pictures to upload. The latter work presents a set of fixed pictures the user needs to choose to automatically build her profile (a faster strategy), presenting a combination of travel profiles she is most compatible with. The user then needs to indicate various attribute preferences to obtain POI recommendations. Although the profile elicitation is easier, it does not explicitly indicate the type of preferred tourist attractions and therefore is not so detailed as a profile that uses the users’personality to predict more refined tourist preferences, as we propose. Instead, it proposes a combination of tourist typologies, which has a broader range of preferences which do not apply to all kinds of tourists. In this work, we propose a solution to obtain a more detailed profile with distinction on the preferred tourism categories using the users’personality, that uses the personality models previously suggested (Alves, Martins, Saraiva, et al., 2023), instead of a broader tourist typology. The existing GRS try to find a satisfactory recommendation to fit all the group members, but most of them are more suitable for small groups, need a great amount of data to build the first recommendations or need many interactions and negotiation between the users to obtain the final recommendation. To use the tourists opinions and negotiation is a very complex task in large and casual groups (e.g., 20 +members), like excursion groups organized by travel agencies, ending up with unsatisfied members when the groups are large and heterogeneous, because, as evidenced by Jameson, Baldes, and Kleinbauer (2003), they were persuaded to choose an item they did not want or because they felt the need to satisfy the other group members choosing unwanted items. We propose to solve those problems (open issues (c), (e), (f) and (g)) by, instead of trying to find a list of POI to satisfy all the group members at once, divide the main group into subgroups of similar personalities, and therefore preferences, automatically assigning to each subgroup a preliminary list of POI to visit based on the members’personality, alleviating or even eliminating the posterior negotiation phase, as the group was divided into smaller subgroups and can use ratings to accept or discard POI in the recommendation list without the influence of the other subgroup’s members (more details in Sections 2.3 and 3.4). 2.2. Agent-based Recommender systems To help with limitations found in Recommender Systems, some authors are also proposing the use of Multi-Agent Systems (MAS) to automatically learn the user’s profile and proactively make recommendations (Alves et al., 2022; Batet, Moreno, S´ anchez, Isern, &Valls, 2012; Borr` as, Moreno, &Valls, 2014). The autonomy, elasticity, distribution, isolated state, and loose coupling of MAS (Collier, O’Neill, Lillis, &O’Hare, 2019; Wooldridge &Jennings, 1995) allow the creation of distributed systems (Batet et al., 2012), characteristics similar to the ones of Microservices, permitting the deployment of intelligent agents as microservices, the Multi-Agent Microservices (MAMS) architecture (Collier et al., 2019). This architecture can take us further if implemented in a GRS, by creating easily accessible agents and a faster system. The use of MAS in Recommender Systems is expanding in the most varied domains (Carneiro et al., 2020; Neto, Morais, Gonçalves, & Coelho, 2022a,b), but only few, and complex approaches, related to tourism can be found. Bekkerman, Kraus, and Ricci (2006) used an individual and self-learning case-based recommender system for tourism assets, Trip@dvice, to develop a GRS. The GRS uses automated negotiation agents, that represent the group members, which cooperatively negotiate and use the individual recommendations and user preference P. Alves et al. Expert Systems With Applications 255 (2024) 124812 3 models to provide group recommendations. Lorenzi, Loh, and Abel (2011, 2012) designed a Multi-Agent RS for travel agencies to help users find personalized travel packages. Using the Distributed Artificial Intelligence paradigm, different agents can be used simultaneously to respond to requests asynchronously. The requests are divided into partial tasks and handled by distinct agents, who search information in their local knowledge base following the Case-based Reasoning approach, or if the needed information is not found, in the community by cooperating with other agents. The final recommendation is then presented to the user, who evaluates it, contributing to the agents’confidence degree. The RS proposed by Batet et al. (2012) provides personalized recommendations of cultural and leisure activities, to the tourists visiting a city, also using a MAS. The architecture consists of different types of agents, where each tourist is represented by a User Agent, the tourism activities by Activity Agents, a Broker Agent that allows the communication between the User and Activity Agents, and the Recommender Agent, which keeps a repository with all the user profiles, with contextaware dynamic updates, providing recommendations based on collaborative and content-based recommendation techniques. Ravi, Devarajan, Sangaiah, Wang, and Subramaniyaswamy (2021) proposed a Multi-Agent cloud-based framework for recommending personalized POI. The framework implicitly gathers the user’s travel behavior and activity from social networks, and six agents are responsible for detecting the changes in the user’s behavior and preferences, to dynamically update the user’s profile, aided by a reference ontology. The recommendations list is generated by using a multi-level collaborative filtering algorithm and the ontology created by the MAS. Although an interesting work, there are no details on how the MAS actually works. Neto, Morais, Gonçalves, and Coelho (2022a,b) proposed a contextbased Multi-Agent RS to help evacuate the occupants of buildings in case of fire, by providing real-time information on the faster evacuation paths using the users’context data obtained from the surrounding Internet of Things objects. The use of Agent-based Modeling is also a technique based in MAS that can be useful in the tourism context to simulate different complex scenarios and determine changes caused by the tourists’actions or choices by using different machine-learning techniques (Wallinger, Grundner, Majic, &Lampoltshammer, 2023). In this work, we use a MAMS to facilitate the recommendation process and the access to the agents’knowledge through REST endpoints. We tried to simplify the implementation and communication process by only modeling two types of agents, one to represent the tourists, and another to represent the (sub)groups. 2.3. Solving conflicting preferences by clustering tourists with similar personalities The interaction and relationship between group members with different personalities, and consequently, different tourist preferences are susceptible to conflicts, as the members exchange their divergent opinions and to negotiate and reach a consensus might be a very difficult task, especially in large occasional groups of tourists. This is aggravated by the well-known contagion phenomenon, which can easily spread within the group and negatively or positively “contaminate”the other members (Delic &Masthoff, 2018; Deli´ c et al., 2020), leading to recommendations that can falsely satisfy the group members. Also, the social influence and contagion phenomenon in choice processes can lead to more unsatisfactory recommendations than recommendations based on well-defined facts and rules. For instance, considering the ThomasKilmann conflict resolution styles (Thomas, 2008), a very dominating tourist can convince other tourists to visit a POI he likes more than the other tourists. He becomes satisfied but the other tourists don’t. This result was demonstrated on a study conducted by Delic, Neidhardt, Rook, Werthner, and Zanker (2017), who found passive participants were the most unsatisfied individuals and that mixed types led to more satisfactory results. They also found that more open to experience, less neurotic and agreeable persons were the most satisfied with the recommendations. Pei, Gao, Yan, Zhou, and Wu (2024) proposed to solve conflicting preferences by using opinion dynamics, trust dynamics, and scorings. It is well evidenced that how an individual copes with a conflicting situation can influence the group decisions (Carneiro, Martinho, Marreiros, &Novais, 2018; Thomas, 2008), as well as the emotions, social relationships, and social identity (Deli´ c et al., 2020). However, we believe this is not the main focus of intelligent systems, i.e., one of the advantages of using intelligent systems is to overcome certain human limitations and improve their quality of life. Considering all that, we believe dividing the main excursion group into subgroups of similar interests, and providing preliminary recommendations based on the users’profile and ratings given to POI visited in the past, can provide a more accurate, fair and satisfactory recommendation list of POI, than a list that was constructed based on the opinions exchanged by the tourists during the process with the social influence aforementioned, i.e., we believe the social influence can negatively affect the (sub)groups satisfaction and travel experience. Although it is demonstrated that decisions are better made in groups (Michaelsen, Watson, &Black, 1989; Shaw, 1932), in the case of choosing a tourist attraction to visit, the influence of others can mislead a tourist to visit places she will not like. Although many preference aggregation strategies exist (Deli´ cet al., 2020; Ismailoglu, 2022; Massimo &Ricci, 2021; Masthoff, 2015; Najafian, Herzog, Qiu, Inel, &Tintarev, 2020), we believe that clustering tourists based in their raw personality instead of different types of preferences/ ratings can lead to faster recommendations, a consensus without much or any negotiations, and more satisfied tourists. As well pointed by Forsyth (2018), group decisions seldom aggregate the opinion of the majority, and certain aggregation strategies can be more suitable for certain groups than others (Masthoff, 2015; Najafian et al., 2020), like the case of occasional large tourist groups. Hence, if a tourist is assigned to a cluster with tourists of similar personalities, the tourist attractions preference will also be similar, which reduces the main group heterogeinity and can avoid conflicts and biased opinions in the choice process, leading to more precise and personalized group recommendations. With this clustering strategy we propose to solve open issues (c) and (e), and (f) by providing recommendations considering each (sub)group’s needs (disabilities and fears/ phobias). More details can be found in Section 3.2. Several models that use clustering to improve group recommendations can be found in literature. Ntoutsi, Stefanidis, Nørvåg, and Kriegel (2012) propose gRecs, that pre-partitions users into clusters of users with similar preferences using hierarchical agglomerative clustering. Using collaborative filtering, they predict the individual ratings (preferences) for the unrated items for each user in the respective cluster and use them to calculate the aggregated ratings for the group. The recommendations for a user are then based on the respective cluster members’ preferences, eliminating the need for searching similar users in the database, suggesting the top-kitems along with explanations, as explanations are important to help the users understand the recommendations (Barkan, Bogina, Gurevitch, Asher, &Koenigstein, 2024; De Campos, Fern´ andez-Luna, &Huete, 2024; Najafian et al., 2020). Boratto and Carta (2015) explored group recommendation approaches that use clustering to automatically detect groups of users with similar preferences. They argue the best approach to predict group preferences for an item is to use an average of the individual preferences, and that using all the predictions generated for individuals is more accurate than other approaches. They also give some advice to help in the development of GRS that automatically detect groups. Using the MovieLens 1 M and 10 M datasets, Bi, Qu, Wang, and Shen (2017) propose a group-specific singular value decomposition method that extends the singular value decomposition (SVD) model by incorporating betweensubject dependency, taking advantage of missing information from users or items (movies) to cluster users according to the quantiles of the ratings they have provided, and the items according to the release dates, assuming users or items from the same cluster share the same group effects. They evaluate the efficacy of the proposed approach by P. Alves et al. Expert Systems With Applications 255 (2024) 124812 4 comparing its performance against four competing matrix factorization models under the regularized SVD, finding their approach demonstrates superior performance across all configurations. Nozari and Koohi (2020) present Influence-Based Group Recommendation (IBGR), that uses Fuzzy C-means and the Pearson Correlation Coefficient to cluster users of compatible interests. Their main focus is how the group members influence each other by using trust and similarity to estimate new ratings. IBGR modifies the ratings of its members by taking into account their influence, i.e., the adjustment of a member’s rating for an item is determined by subtracting their rating from the ratings of other members, multiplied by the respective influence weight assigned to each member. The average aggregation strategy is then used to calculate the group’s rating for each item, considering the influencedbased ratings. The top-kitems are then recommended. Although interesting, this proposal is complex and goes against our objective of avoiding the contagion phenomenon associated with groups. An online learning algorithm, cluster-based bandits, is proposed by Shams, Anderson, and Leith (2021) to help with the cold-start problem. The algorithm is capable of rapidly learning the other users’preferences using only 10 to 12 rated items to reliably distinguish between 16 and 32 user groups, respectively. To cluster the users into groups, they use training data to predict the ratings’mean and variance for a certain item for each group, which is tested with different numbers of clusters. The increased demand for recommendations and the items available has caught the attention of researchers, which are developing algorithms capable of dealing with high-throughput data streams and perform streaming recommendations, dynamically evolve with their content and/ or generating online the used models (Chandramouli, Levandoski, Eldawy, &Mokbel, 2011; He et al., 2023; Liu, Zhao, Wang, Liu, &Tang, 2020; Lourenço, Meira, &Marreiros, 2023; Silva et al., 2013). Although interesting and useful approaches, Grouplanner is targeted for tourists and not for high-throughput data stream services, as the number of users will not increase indefinitely, and it was not programmed to handle transient large volume data (Silva et al., 2013). That is not the purpose of the proposed GRS. However, the proposed algorithm is dynamic, by assigning a new user and eventually updating existing ones in real time, i.e., it processes the information as soon as it is received. Also, due to impositions of the Grouplanner application, the users’profile is saved in the database, which is not suitable for a streaming algorithm that uses fixed memory and single scan, as it implies forgetting part of past data, data that is crucial to perform the clustering and apply the association rules. Although various clustering algorithms exist and some are being used in recommender systems (Ackermann et al., 2012; Aggarwal, Philip, Han, &Wang, 2003; Bengali &Veldt, 2023; Ezugwu et al., 2022; Mai, Assent, &Storgaard, 2016; Massimo &Ricci, 2021; Saxena et al., 2017), they need previous interaction data and/or ratings given by other users to provide the first recommendations. Also, the number of clusters needs to be known a priori. We believe this is not the best approach, as it forces us to know all the referred information in advance, not actually solving the cold-start problem. Besides, the fact that someone gives the same rating as another user does not mean they rated it for having the same reason. The underlying reason is crucial to understand if the ratings given are similar, so the clusters are the most accurate and representative of the whole group as possible. For example, a user might have rated a beach with 5 stars because she appreciates its natural beauty and another user because he just likes to swim in the sea and the water of that beach is perfect for that. Also, it is common to find users in the same cluster with different ratings for the same item, which adds noise to the ratings (Shams et al., 2021). Therefore, if instead we aggregate users based on their personality similarity, they will have similar preferences and ratings for the same reasons, increasing the clustering quality. To accomplish that and solve the cold-start problem, we modified the k-means algorithm, creating d-means, an algorithm that incrementally creates/modifies clusters in real-time as the users register in the application, taking advantage of the users’personality, without needing to define the number of clusters a priori, previous interaction data and ratings (see Section 3.2). 3. Methodology 3.1. The prototype architecture The Grouplanner prototype is a mobile application developed for Android platforms using Android Studio, Visual Studio, Java and the. NET framework (Alves, Martins, Novais, et al., 2023). As can be observed in Fig. 1, Grouplanner is composed of five MS, deployed in a free Microsoft Azure server, and the communication between them is performed by sending requests to HTTP REST endpoints. The MS have the following features: User Management MS (UMMS) –responsible for managing and persisting the registered users’data (demographic data, personality, disabilities (physical impairments, deafness, blindness, heart conditions), fears/phobias (heights (acrophobia), claustrophobia, crowded spaces (agoraphobia)), tourist attractions preference, travel-related preferences &concerns, (un)visited POI, rated POI and other profile related information); Multi-Agent MS (MAMS) –composed of agents created in the MAMS environment every time a new tourist user registers in the application or a new excursion group is created. Every user is represented by an intelligent agent modeled with the respective profile. The MAMS is responsible for creating dynamic clusters of similar personality tourists and in the case of excursion groups, dividing the main group into subgroups based on those clusters and on the excursion’s tourists’ personality; Recommendation Engine MS (REMS) –responsible for finding the best POI to recommend to the tourists in a POI ontology (Silva, 2023), according to their tourist attractions’preference, travel-related preferences &concerns, disabilities, fears/phobias, and the restrictions found by the MAMS; POI MS (POIMS) –Contains all the POI to be suggested, persisted in a repository, obtained from the TripAdvisor API (at the moment only POI from the north of Portugal are available in the app). The POI microservice is responsible for managing all the POI. Communicates with the Google Places API to retrieve the information needed for a user to register a newly visited POI or to show updated information on the POI presented in the recommendation lists. Social Network MS (SNMS) –manages the groups and subgroups’ information and chats, essential for future consultations within the app and presentation of their information in the user interface. All microservices repositories use a PostgreSQL database. Fig. 1. Grouplanner Microservices Architecture (first published in Alves et al. (2022)). P. Alves et al. Expert Systems With Applications 255 (2024) 124812 5 At the moment, the Grouplanner prototype allows the registration of tourist users; their profile update and incremental improvement; to register POI visited in the past, rate them from 1 to 5 stars, and define if they want to visit them again; search for other registered users and view their visited POI history; create excursion groups, define excursion dates, add users to the excursion (registered and/or external users, like accompanying relatives), force pairing with other group members, and chat with the group members; request individual and group recommendations for specific dates. Although they are not used to provide the initial recommendations, the tourists’POI history and respective ratings combined with the tourists’profile are used to improve the recommendation process, i.e., the agents in the MAMS can learn from the tourists’travel experience and profile to reason on rules for the existing clusters, and consequently, to the corresponding tourists. The capability of providing individual recommendations is crucial, as the agents responsible for the subgroups (or main group if no subgroups are formed) can learn with the evaluation given to the suggested POI by the respective tourist, and therefore gain knowledge to help provide better individual and group recommendations. Also, if the tourists are satisfied with the individual recommendations, it means a more solid base can be provided for the groups’recommendations. 3.2. Registration of tourists with dynamic clustering When a tourist user registers in the application, she needs to fill some basic profile information (name,username,email,gender,birthdate,country, city,nationality, fears/phobias, and disabilities) and the personality questionnaire (Big Five Inventory (BFI), 44 items, which assesses in a 5-point Likert scale the respondent’s score on the Goldberg’s five dimensions of personality (Goldberg, 1990)), as can be seen in Fig. 2(a), Fig. 2(b) and Fig. 2(c). It is important to consider that, if the tourist does not respond truthfully to the BFI, the recommendations can be less accurate. In the BFI screen, the users are warned to respond honestly so they can receive relevant recommendations. It is therefore up to the user to respond honestly or not to the personality questions. The user’s profile is then sent via HTTPS POST to the UMMS and MAMS, being the communication protocol detailed in Alves, Martins, Novais, et al. (2023). As is evidenced in literature, completing personality questionnaires can be a tedious task and is subject to the social desirability bias usually associated to self-reporting questionnaires (Pedregon, Farley, Davis, Wood, &Clark, 2012). Although the prototype uses the BFI to determine the tourists’personality, as previously mentioned, we are working on ways of gamifying the application by developing minigames to implicitly detect their personality (Alves et al., 2024), namely the 30 personality traits of the Five Factor Model of personality (Costa &MacCrae, 1992; Digman, 1990). With this approach, we will not only avoid intrusive questionnaires, but also the social desirability bias and untruthful answers, and the need for great amounts of data to determine personality by other implicit methods (Finnerty, Lepri, &Pianesi, 2016; Tkalcic &Chen, 2015), improving at the same time the tourists’ involvement and motivation while using the application (Xu, Tian, Buhalis, Weber, &Zhang, 2016). The normalized personality scores are used by the MAMS to predict the tourist’s preference in 11 categories of tourism and in 4 travelrelated preferences &concerns (see Fig. 2(d), Fig. 2(e), Table 1 and Table 2). The referred items result from an intensive study to determine the relationship between an individual’s raw personality (Big Five) and her preference for tourist attractions and travel-related preferences & concerns, having found they are strongly related, resulting in the proposal of structural equation models that confirm that relationship (Alves, Martins, Saraiva, et al., 2023). Those models are used by the MAMS to calculate a prediction on the preference for the aforementioned travel aspects, bypassing the initial learning phase needed by the existing RS (Dara, Chowdary, &Kumar, 2020), solving the cold-start. As previously mentioned, we adapted the k-means algorithm to aggregate the tourists with similar preferences. If the tourist is the first user registering in the application, a cluster based on his personality scores is created, becoming the centroid, 2 otherwise, his similarity to the center of the existing clusters is calculated and, if the similarity is 0.80 or more (see Equation (2), (configurable, to introduce more or less sensitivity), he is added to the most similar one or to a new cluster if not, constituting the proposed d-means algorithm. To maximize the tourists’ similarity inside a cluster, the formation of the dynamic clusters is solely based on the tourists’personality. This modification of the k-means algorithm eliminates the need to know the number of clusters in advance, as mandatory by the clustering algorithms (Ezugwu et al., 2022), as we already know the characteristics we want to cluster, i.e., the five dimensions of personality. Also, it solves the k-means problems related to noise and outliers, since the tourists are only assigned to a cluster if they are close to its centroid, not being forced to cluster if they are far away. Therefore, the clusters will contain users with highly similar personalities, and consequently, preferences, which will help mitigate the heterogeneity and conflicting preferences associated with casual/large groups. To solve problems related to sparsity (Altman &Krzywinski, 2018), when an excursion group is created, it is divided into subgroups of similar personality, and users that were alone in a cluster will automatically belong to one of those subgroups (see Section 3.4). We chose k-means over the other mainstream clustering algorithms because hierarchical clustering algorithms are more suitable for situations where all objects of interest are in a main cluster that needs to be divided into smaller parts, or the opposite, which is not the case in the current problem. This left us with partitional clustering approaches, which are the contrary of hierarchical clustering, where the objects are assigned to a number of clusters determined a priori, according to a criterion function (Lam &Wunsch, 2014), which is what we needed for the tourists registering in the application, i.e., to group them according to their personality similarity using a similarity measure. k-means was therefore chosen, as it is one of the most popular and simple clustering algorithms (Ezugwu et al., 2022; Saxena et al., 2017), being easier to adapt to the problem in hand. The implemented clustering process is explained next. Definition 1. Let Ti,i∈ {1,2, ..., n}, be a certain tourist i, where nis the number of tourists registered in the application, and pj,j∈ {1,2, ..., m}, be a certain participant j, where mis the number of participants in the conducted simulation (Section 4.3). Definition 2. Let P∈ {O,C,E,A,N},Ti.P∈ [0,1], where Pis the personality of tourist’s Ti, and Ti.Orepresents the tourist’s Openness to Experience score, Ti.Cthe Conscientiousness score, Ti.Ethe Extraversion score, Ti.Athe Agreeableness score, and Ti.Nthe Neuroticism score. Definition 3. Let APTiFj∈ [0,1],i∈ {1,2, ..., n},j∈ {1,2, ..., 11},be the tourist’sTipredicted tourist attraction preference AP for a certain tourism category Fj. Definition 4. Let TCTiFPk∈ [0,1],i∈ {1,2, ..., n},k∈ {1,4}be tourist’s Tipredicted travel-related concerns TC for a certain travel-related concern factor FPk, and TPTiFPk∈ [0,1],k∈ {2,3}, the tourist’sTipredicted travel-related preference for the travel-related preference factor FPk. Definition 5. Let Cq,q∈ {1, ..., n}the cluster to which a tourist Ti belongs to and Cq.centroid = {O,C,E,A,N}the Cqcluster’s centroid personality, where Ois Cqtourists’average openness, Cis Cqtourists’ average conscientiousness, Eis Cqtourists’average extraversion, Ais Cq tourists’average agreeableness and Nis Cqtourists’average neuroticism, all ∈ [0,1]. 2 The clustering process is only executed if the tourist has filled the BFI first and the respective personality scores calculated. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 6 To compare a tourist Tjto a cluster’s centroid, the dissimilarity between them, d(Tj,Cq), where j∈ {1,2, ..., n},d∈[0, 5 √], was calculated, by measuring the Euclidean distance between their personality (5dimensional), as it is considered the most used metric for numerical data (Ezugwu et al., 2022; Saxena et al., 2017): The same formula is applied to measure the distance between two tourists (points). The closer the distance is to zero, the closer the tourists are. As the Euclidean distance measures the dissimilarity between two points, we used the following formula to calculate the similarity between them (Feldman &Sanger, 2007), so the values were easier to compare: sim(Tj,Cq)=1/(1+d(Tj,Cq)),∈[1/(1+ 5 √),1](2) Where sim(Tj,Cq)represents the similarity between tourist Tjand cluster Cqin terms of personality. The closer to 1 the more similar the tourist and the cluster’s centroid are. In our algorithm, we consider a tourist’s personality is similar to a cluster’s personality (or other tourist’s personality) if sim(Tj,Cq)≥0.80 (this value can be changed to increase or decrease the desired similarity and consequently, the clustering sensitivity, being this the most important input parameter that needs to be configured to run the algorithm. All the other application parameters and default values can be consulted in Appendix C Table C2. The dynamic clustering process is presented in Algorithm 1. After running Algorithm 1, a tourist that once belonged to a cluster Cxcan now belong to a cluster Cy, as the clusters’centroids are recalculated and the tourists’clusters updated. Fig. 2. Some Grouplanner app screens: (a) new user registration form; (b) personality inventory (BFI); (c) the tourist’s personality scores after filling the BFI; (d) the tourist’s predicted attraction preference; and (e) the tourist’s predicted travel-related preferences &concerns. Table 1 Description of the tourism categories available for recommendation, adapted from Alves, Martins, Saraiva, et al. (2023). Tourism categories Description Adrenaline Activities (F1) Adrenaline promoting activities (hard adventure), like extreme sports, such as aquatic sports (e.g., sailing, canoeing), motor sports (e.g., rally, karting, motocross), under water activities (diving, snorkeling), climbing, bungee jumping, air sports (e.g., skydiving, parachute jumping, gliding), ski. Wild Nature Activities (F2) Soft adventure activities related to wild nature, such as to appreciate natural landscapes, hiking, mountaineering, visit nature or wildlife reserves, walk in forests/woods, walk along rivers or seacoasts, visit mountain areas or gorges. Party, Music &Nightlife (F3) Events related to celebrations, listening to music, watching movies and nightlife, like attending discos or night clubs, live music bars/places, thematic parades (military, electronic, etc.), film festivals, music festivals, concerts, dance/ballet festivals, balls. Sun, Water &Sand (F4) Activities related to relaxing or swimming at a swimming pool or the beach, sunbathing, take vacations on an island. Museums, Boat Trips & Viewpoints (F5) Activities involving the visit to museums (historical or scientific), the interest in nature or urban viewpoints, or boat trips to know the coast or to appreciate its historical value. Theme &Animal Parks (F6) This category includes activities usually done in family, like going to water parks, funfairs, a zoo or theme parks (like Isla Magica or Disneyland Paris). Cultural Heritage (F7) Activities associated with knowing the cultural heritage of a country, like visiting monuments (e.g., castles, monasteries, churches, cathedrals) and historical cities or villages. Sports &Games (F8) This category is related to hunting, fishing or assisting to a sporting competition (e.g., football game). Gastronomy Events (F9) Events related to gastronomic experiences, like gastronomy tours or festivals involving food (typical and/or gourmet dishes) and wine tasting. Health &Well-being (F10) Activities to promote the health &well-being, such as going to a SPA, beauty centers, or do health and wellness treatments (e.g., mineral water resorts). Natural Phenomena (F11) Assist to natural events like observing natural phenomena (e.g., volcanic eruptions, northern lights, waterfalls), visit caves, caverns, or volcanoes. d(Tj,Cq)= (Tj.O−Cq.O)2+(Tj.C−Cq.C)2+(Tj.E−Cq.E)2+(Tj.A−Cq.A)2+(Tj.N−Cq.N)2 √(1) P. Alves et al. Expert Systems With Applications 255 (2024) 124812 7 Algorithm 1 Dynamic personality-based clustering, d-means Input: New tourist’sTipersonality Output: Cluster Cqcontaining tourist Tiand all clusters with centroids updated 1://Assignment of the registering tourist to the initial cluster 2: Obtain tourist’sTipersonality 3: maxSimilarity←0,Cc←null 4: if list〈Cluster〉is empty 5: create cluster C1and set Tias centroid (C1.centroid = {Ti.O,Ti.C,Ti.E,Ti.A,Ti.N}) 6: add C1to list〈Cluster〉 7: else 8: for each cluster Cqin list〈Cluster〉do 9: calculate normalized Euclidean distance sim(Ti,Cq.centroid)between Tiand Cqcentroid 10: if sim(Ti,Cq.centroid) ≥ 0.80 and sim(Ti,Cq.centroid)>maxSimilarity then 11: set Cqas closest cluster Cc,maxSimilarity←sim(Ti,Cq.centroid) 12: end if 13: end for 14: if Cc∕= null then//update cluster Cccentroid 15: add Tito Cc 16: Cc.O=∑Cc.size j=1Tj.O Cc.size ,Cc.C=∑Cc.size j=1Tj.C Cc.size ,Cc.E=∑Cc.size j=1Tj.E Cc.size ,Cc.A=∑Cc.size j=1Tj.A Cc.size ,Cc. N=∑Cc.size j=1Tj.N Cc.size 17: else//Tiwas not added to a cluster 18: create new cluster Cxand set Tias centroid (Cx.centroid = {Ti.O,Ti.C,Ti.E,Ti.A,Ti.N}) 19: add Cxto list〈Cluster〉 20: end if 21: end if //Recalculation of the clusters’centroids (adapted from k-means algorithm) 22: for each tourist Tjin list <Tourist >do 23: maxSimilarity←0,Cc←null 24: for each cluster Cqin list〈Cluster〉do 25: calculate normalized Euclidean distance sim(Tj,Cq.centroid)between Tj and Cqcentroid 26: if sim(Tj,Cq.centroid)≥0.80 and sim(Tj,Cq.centroid)>maxSimilarity then 27: set Cqas closest cluster Cc,maxSimilarity←sim(Tj,Cq.centroid) 28: end if 29: end for 30: if Cc∕= null then 31: add Tjto Cc 32: else//Tjwas not added to an existing cluster 33: create new cluster Cxand set Tjas centroid (Cx.centroid ={Tj.O,Tj.C,Tj.E,Tj.A,Tj.N}) 34: add Cxto list〈Cluster〉 37: end if 35: end for 36: for each cluster Cqin list〈Cluster〉do 37: recalculate cluster Cqcentroid (same as step 16) 38: end for 39: repeat steps 22 to 38 until the existing clusters’centroids no longer change k-means (Abernathy &Celebi, 2022; Arthur &Vassilvitskii, 2007) is NPhard and the time complexity is O(I×(n×k×d)), where nis the number of points, kthe number of clusters, dthe number of dimensions of each points, and Ithe number of iterations performed. In d-means, the main loop iterates until no changes in the centroids are detected, which depends on the number of Iiterations. The inside loops iterate over all users and clusters to determine their distance, resulting in a time complexity of O(n×k×d)for each main loop iteration. The centroid calculation adds a complexity of O(n×k)also for each main loop iteration. This results in a time complexity of O(I× (n×k×d+n+k)). This personality-based dynamic clustering is an innovative approach, used to the best of our knowledge for the first time in GRS, accurately assigning the tourists to clusters of other similar tourists in real-time as they register in the application, eliminating the need for previous interaction data and/or ratings as in the existing RS, solving the cold-start problem. This led to an alternative algorithm to the previously proposed (Alves, Martins, Saraiva, et al., 2023). 3.3. Determining rules to improve recommendations To refine the predicted recommendations, using association rules, namely the Apriori algorithm with at least a 95% confidence and support intervals, the MAMS searches for positive and negative rules for each cluster, considering the tourists’profile (demographic data, weather preference, disabilities, fears/phobias (more details in Table 3)) and ratings given to (un)visited POI. 3 This process creates a list of positive and negative rules for each cluster. For example, let’s imagine several tourists with 45+years old from cluster C1rated positively museums (4 or 5 stars). A positive rule could be: “Tourists with 45+ years old like to visit museums”. A negative rule could be: “Tourists who have children do not like to visit Tower of Cl´ erigos”. This algorithm is programmed to run as a job during late night, once a day, to update the clusters’rules (parameterizable). The rules are searched for in a cluster if it has at least 50 members (configurable), otherwise, the rules are searched among all the tourists in the app. Table 3 Users’attributes, and respective classes, used by the MAMS to determine the association rules list for each cluster. Attribute Classes Age ≤23, between 24 and 34, between 35 and 45, ≥46 years old (25 % bins) Gender Female, Male Marital status Single, In a relationship, Married, Other Children No, Yes Educational level High school, Degree, Master, Ph.D., Other Formation area Engineering &Technology, Medical &Health Sciences, Natural Sciences, Exact Sciences, Humanities, Social Sciences, Other, None Professional situation Employed, Student, Working student, Self-employed, Other Live with Parents, Partner &Children, Partner, Friends, Alone, Children, Other Liquid income Not applicable, <650, between 650 and 1000, between 1001 and 2000, >2000 € Professes religion Yes, No but believe in a superior being, Atheist, Agnostic, Other Travel companions Relatives, Partner, Partner &Children, Friends, Other Disabilities Physical, Sensorial, Other, None Fears/phobias None, Heights, Claustrophobia, Agoraphobia, Other Weather preferences No preference, Cold weather, Warm weather, Hot weather Ratings given to (un) visited POI Don’t like, Neutral, Like Table 2 Description of the travel-related preferences &concerns considered in the recommendations, adapted from Alves, Martins, Saraiva, et al. (2023). Travel-related preferences &concerns Description Previsibility and Safety (FP1) Travel-related concerns, like to feel comfort, have mobile phone signal, fear of getting lost, fear of getting ill or having an accident, not to take risks related to the physical integrity. Cultural &Learning Experiences (FP2) Travel-related preferences, like studying the destination before traveling, consider the destination’s cultural offer by including a cultural/learning component with a visit to iconic monuments or even visit as many attractions as possible. Uniqueness &Exoticness (FP3) Travel preferences related to visiting “exotic”points of interest, uncommon places or observe peculiar things. Familiarity (FP4) Concerns related to traveling in a familiar environment, with known acquaintances (family, friends, etc.), or not to travel in a group with strangers. 3 As the ratings in the database are in a 5-point Likert scale, values between 1 and 2 are categorized into “1-Don’t like”, 3 to “2-neutral”, and 4 and 5 to “3Like”. Category 1 is considered negative, and Categories 2 and 3 positive. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 8 To help kickstart this process and bypass the rules cold-start, as the number of users registered in the prototype was low, we registered the 1035 participants in the dataset 4 from our previous study (Alves, Martins, Saraiva, et al., 2023) in the database, so it could be considered by the MAMS and the clustering process, as it has the participants personality and their tourist preferences. 3.4. Forming subgroups to tackle heterogeinity and conflicting preferences To help with the heterogeinity and conflicting preferences associated to occasional/large groups of tourists, as the tourists are assigned to a cluster of other similar-personality tourists when they register in the app, that clustering is used by the MAMS as an advantage to divide the excursion groups into subgroups of also similar-personality tourists. This means that tourists from an excursion group who belong to the same cluster are gathered in a subgroup, which will drastically reduce the main group’s heterogeneity and conflicting preferences, as they will have similar preferences due to their personality. This can be advantageous, as it is evidenced that smaller groups have higher satisfactory travel experiences and perceived value than larger ones (Su, Cheng, & Huang, 2021). This division is only performed if the subgroups can have at least three members and at least two subgroups can be formed. 5 The process is detailed next and in Algorithm 2. The creation of similar personality subgroups can be the solution for sparsity issues (sparse users), related to the curse of dimensionality (Altman &Krzywinski, 2018), as users that are alone in a cluster will be forced to belong to a subgroup with the most similar personality. When a group owner requests a recommendation for an excursion group, Fig. 4 and Fig. 5, if at least three tourists in the group belong to the same cluster, a subgroup, CqSg, where Cqrefers to the originating cluster, with the respective tourist agents is created as well as a subgroup Travel Agent, TrvAg-<GroupID>sub < ClusterID>, responsible for the subgroup’s tourist agents interaction. If there are tourists from clusters that cannot generate a subgroup, i.e., that would create a subgroup of only 1 or 2 members, they are individually compared to the clusters’ centroid that originated the already formed subgroups, using the normalized Euclidean distance (Equation (2), and included in the most similar subgroup (Fig. 3). It is important to notice that neither the subgroup’s centroid nor the originating cluster’s centroid are updated at this stage. All this guarantees that only the most similar tourists, and consequently with similar tourist preferences, are in the same subgroup and that the corresponding subgroup’s and cluster’s centroid is not “damaged”by the posteriorly added tourists’personality. If at least two subgroups could not be formed, the main group remains intact, and the recommendations are considered for the full group. Algorithm 2 Division of the excursion group into subgroups of similar-personality tourists Input: Excursion group Output: Subgroups of similar-personality tourists or the main excursion group 1: ExcursionGroupexcursionGroup; 2: 3: //to form at least 2 subgroups of 3 tourists or more, the excursion group needs to have 6 or more members 4: if excursionGroup.size ≥6then 5: 6: HashtableclustersHash =newHashtable(); 7: 8: for each Tiin excursionGroup.tourists do//count the number of tourists from same cluster 9: if Ti.Cqnot in clustersHash then 10: clustersHash.add(Ti.Cq,1);//Cluster Cqis the key 11: else 12: clustersHash[Ti.Cq]+ = 1; 13: end if 14: end for 15: 16: count number of clusters with at least 3 tourists 17: 18: if count ≥2then//guarantees there will be at least 2 subgroups 19: for each Cqin clustersHash do 20: if clustersHash [Cq]>2then//form subgroups of at least 3 similarpersonality tourists 22: create subgroup CqSg with the corresponding tourists//Cqrefers to the originating cluster 22: create travel agent TrvAg-<ExcursionGroupID>sub Cq.id 23: add TrvAg-<ExcursionGroupID>sub Cq.id to excursionGroup.subgroups[CqSg]; 24: add subgroup CqSg to excursion.subgroups list; 25: end if 26: end for 27: 28: for each Cqin clustersHash do 29: if clustersHash[Cq]<3then//if only 1 or 2 tourists are from the same cluster 30: for each Tiin Cqdo 31: int maxSimilarity ←0; 32: for each CqSg in excursionGroup.subgroups do 33: calculate normalized Euclidean distance sim(Ti,CqSg.centroid)between Tiand CqSg centroid 34: if sim(Ti,CqSg.centroid)>maxSimilarity then 35: set CqSg as closest subgroup CqSgclosest 36: maxSimilarity←sim(Ti,CqSg.centroid); 37: end if 38: end for 39: if CqSgclosest!=null then 40: add Tito CqSgclosest//add tourist Tito closest subgroup 41: update excursionGroup.subgroups accordingly 42: end if 43: end for 44: end if 45: end for 46: else 47: keep main excursion group intact 48: end if 49: else 50: keep main excursion group intact 51: end if If a tourist is traveling with his family and/or other companions, and they are registered in the application, he can force the application to pair with their respective users, so when subgroups are formed, they will stay together. If they are not registered, the tourist can add them to the excursion group as external members and only fill in the profile mandatory information. It is important to notice that the subgroup’s Fig. 3. Representation of the main excursion division into subgroups with at least 3 tourists from the same cluster. 4 The dataset is composed of participants mostly from Portugal and between 15 and 68 years old (70% between 23 and 55), being 74% female, 31% from the “Engineering &Technology”area, 26% from “Social Sciences”, and 43% from other areas. 69% have higher education, 61% are employed workers and 35% students. 60% are in some sort of a relationship and 42% have children. 69% have some sort of fear/phobia (mainly heights). More details can be found at https://www.gecad.isep.ipp.pt/grouplanner/dissemination.html. 5 These decisions were made because, for instance, if there is an impasse in a decision, the third person can break a tie. The presence of a third person can also help mediate conflicts. More persons can also help choose more diverse places to visit. The group dynamics can be richer with three persons than with only 2. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 9 Table 7 Personality, predicted tourist attraction preference and predicted travel-related preferences &concerns of the participants in the clusters formed, grouped by the corresponding subgroup (CqSg). User # CqSg Cluster# O C E A N F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11 FP1 FP2 FP3 FP4 1 11 C21 0.95 0.83 0.50 0.58 0.69 0.40 0.48 0.44 0.40 0.54 0.42 0.58 0.39 0.42 0.48 0.44 0.68 0.63 0.63 0.61 4 11 C11 0.75 0.81 0.28 0.47 0.63 0.25 0.35 0.32 0.30 0.45 0.33 0.47 0.32 0.31 0.39 0.34 0.54 0.63 0.46 0.64 8 11 C11 0.78 0.75 0.34 0.58 0.53 0.31 0.44 0.34 0.33 0.52 0.34 0.58 0.33 0.37 0.39 0.43 0.59 0.56 0.52 0.55 13 11 C11 0.78 0.61 0.25 0.47 0.53 0.29 0.38 0.30 0.27 0.49 0.31 0.47 0.33 0.31 0.34 0.39 0.53 0.51 0.45 0.56 26 7 C4 0.63 0.86 1.00 0.69 0.72 0.73 0.69 0.80 0.82 0.57 0.76 0.69 0.67 0.76 0.78 0.62 0.73 0.71 0.81 0.58 28 7 C7 0.83 0.67 0.88 0.75 0.63 0.71 0.72 0.70 0.68 0.66 0.66 0.75 0.59 0.70 0.66 0.67 0.77 0.56 0.83 0.47 34 7 C7 0.88 0.72 0.78 0.89 0.56 0.63 0.74 0.61 0.60 0.73 0.57 0.89 0.49 0.66 0.57 0.69 0.82 0.54 0.84 0.40 5 3 C3 0.85 0.61 0.53 0.67 0.63 0.49 0.57 0.48 0.44 0.62 0.45 0.67 0.42 0.49 0.47 0.56 0.69 0.53 0.65 0.48 6 3 C3 0.73 0.64 0.63 0.69 0.59 0.48 0.58 0.41 0.34 0.70 0.38 0.67 0.37 0.44 0.36 0.62 0.68 0.55 0.67 0.48 7 3 C17 0.70 0.47 0.78 0.56 0.66 0.70 0.63 0.68 0.65 0.58 0.65 0.56 0.62 0.64 0.64 0.61 0.65 0.52 0.69 0.49 14 3 C6 0.68 0.61 0.59 0.61 0.41 0.53 0.56 0.52 0.51 0.55 0.50 0.61 0.49 0.54 0.51 0.54 0.60 0.47 0.62 0.47 22 3 C6 0.65 0.69 0.59 0.64 0.50 0.50 0.56 0.52 0.52 0.56 0.51 0.64 0.48 0.54 0.52 0.53 0.61 0.55 0.62 0.50 25 3 C3 0.85 0.50 0.50 0.78 0.78 0.50 0.63 0.48 0.44 0.73 0.45 0.78 0.39 0.51 0.45 0.63 0.76 0.55 0.68 0.42 2 14 C10 0.93 0.53 0.47 0.75 0.25 0.47 0.60 0.40 0.35 0.66 0.37 0.75 0.37 0.48 0.37 0.60 0.68 0.33 0.67 0.35 3 14 C1 0.50 0.47 0.81 0.64 0.16 0.72 0.68 0.66 0.68 0.56 0.64 0.64 0.64 0.70 0.61 0.65 0.56 0.34 0.68 0.36 9 14 C8 0.55 0.53 0.81 0.58 0.44 0.70 0.64 0.68 0.69 0.54 0.66 0.58 0.64 0.67 0.65 0.61 0.59 0.48 0.67 0.46 23 14 C17 0.63 0.36 0.78 0.75 0.63 0.74 0.74 0.69 0.66 0.71 0.64 0.75 0.59 0.70 0.61 0.73 0.71 0.48 0.73 0.36 31 14 C14 0.70 0.72 0.78 0.86 0.34 0.63 0.73 0.61 0.62 0.68 0.57 0.86 0.52 0.68 0.57 0.68 0.73 0.48 0.79 0.38 35 14 C14 0.68 0.53 0.75 0.81 0.44 0.66 0.73 0.62 0.62 0.69 0.59 0.81 0.54 0.68 0.57 0.70 0.71 0.45 0.75 0.35 36 14 C16 0.85 0.44 0.72 0.67 0.47 0.67 0.67 0.60 0.56 0.64 0.56 0.67 0.55 0.61 0.55 0.65 0.70 0.40 0.73 0.39 11 9 C9 0.58 0.86 0.47 0.69 0.44 0.37 0.51 0.42 0.45 0.54 0.43 0.69 0.39 0.48 0.46 0.48 0.59 0.60 0.57 0.52 15 9 C9 0.68 0.75 0.47 0.64 0.41 0.40 0.51 0.42 0.43 0.54 0.42 0.64 0.40 0.46 0.45 0.49 0.59 0.53 0.58 0.50 19 9 C9 0.70 0.83 0.47 0.69 0.53 0.38 0.52 0.42 0.43 0.57 0.42 0.69 0.38 0.46 0.46 0.49 0.64 0.61 0.60 0.53 20 9 C19 0.90 0.81 0.59 0.72 0.38 0.47 0.58 0.47 0.45 0.59 0.44 0.72 0.42 0.51 0.48 0.54 0.70 0.50 0.71 0.48 29 9 C18 0.75 0.97 0.72 0.61 0.09 0.51 0.54 0.52 0.55 0.45 0.52 0.61 0.51 0.56 0.56 0.48 0.59 0.47 0.69 0.53 32 9 C9 0.55 0.89 0.44 0.58 0.38 0.34 0.45 0.40 0.44 0.46 0.41 0.58 0.40 0.44 0.46 0.41 0.52 0.59 0.51 0.57 33 9 C9 0.52 0.89 0.63 0.67 0.34 0.47 0.55 0.52 0.57 0.51 0.52 0.67 0.48 0.56 0.55 0.50 0.58 0.58 0.62 0.52 10 15 C12 0.5 0.36 0.19 0.44 0.75 0.33 0.39 0.34 0.31 0.52 0.35 0.44 0.36 0.33 0.36 0.43 0.46 0.55 0.35 0.53 12 15 C15 0.7 0.50 0.34 0.61 0.69 0.39 0.50 0.39 0.36 0.60 0.38 0.61 0.37 0.41 0.40 0.51 0.60 0.54 0.52 0.48 17 15 C15 0.7 0.53 0.50 0.56 0.66 0.49 0.52 0.49 0.46 0.56 0.48 0.56 0.46 0.48 0.49 0.52 0.60 0.54 0.57 0.51 16 5 C5 0.50 0.56 0.31 0.72 0.47 0.35 0.53 0.36 0.37 0.62 0.36 0.72 0.34 0.44 0.36 0.54 0.56 0.51 0.49 0.41 18 5 C13 0.75 0.42 0.22 0.83 0.41 0.33 0.58 0.28 0.24 0.73 0.26 0.83 0.25 0.40 0.25 0.60 0.65 0.38 0.55 0.30 21 5 C2 0.60 0.33 0.44 0.94 0.66 0.51 0.72 0.47 0.45 0.83 0.43 0.94 0.37 0.57 0.40 0.74 0.73 0.48 0.64 0.27 24 5 C20 0.58 0.28 0.47 0.69 0.47 0.55 0.62 0.48 0.46 0.67 0.46 0.69 0.46 0.54 0.43 0.64 0.59 0.31 0.59 0.26 30 5 C5 0.60 0.44 0.28 0.69 0.47 0.37 0.53 0.34 0.33 0.63 0.34 0.69 0.33 0.42 0.33 0.55 0.57 0.45 0.49 0.38 Where O=Openness, C=Conscientiousness, E=Extraversion, A=Agreeableness, N=Neuroticism; F1 =Adrenaline activities, F2 =Wild Nature activities, F3 =Party, Music &Nightlife, F4 =Sun, Water &Sand, F5 = Museums, Boat trips &Viewpoints, F6 =Theme &Animal parks, F7 =Cultural heritage, F8 =Sports &Games, F9 =Gastronomy events, F10 =Health &Well-being, F11 =Natural phenomena; FP1 =Previsibility & Safety, FP2 =Cultural &Learning experiences, FP3 =Uniqueness &Exoticness, FP4 =Familiarity. Personality dimensions scores equal or superior to 0.5 are in bold. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 16 algorithm, for two small sets of users (n=50 and n=150) as the graphics would become very complex for large sets. The results are presented in Fig. 11 and Appendix A Table A2-A7. Observing Fig. 11 graphics, we can perfectly see the clusters formed. For d-means, in Fig. 11(a), clusters C7and C16 have the greatest number of similar members, although some overlapping can be seen mainly between these 2 clusters and between C16 and C14. As an example, user 37 was assigned to cluster C11 be could also belong to C16. However, the Fig. 7. (a) Time taken, in seconds, by each algorithm to cluster 35 to 100 k users, for sim ≥0.80; (b) Time taken, in seconds, by each algorithm to assign a new user when there were already kclusters (x-axis) in the database, from 1000 to 100 k users, for sim ≥0.80. Fig. 8. (a) Time taken, in seconds, by d-means to cluster 35 to 100 k users for different sensitivities; (b) Number of clusters formed by d-means for an increasing number of users in the database for each sensitivity. Fig. 9. (a) Boxplots comparing the d-means clusters compactness for the similarity thresholds of sim ≥0.70 and sim ≥0.80 for 50,000 users. (b) Boxplots comparing the d-means clusters silhouette for the similarity thresholds of sim ≥0.70 and sim ≥0.80 for 50,000 users. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 17 graphics only have 2 dimensions and clusters that seem to overlap may not be overlapping. This is confirmed by the high values obtained for the clusters’compactness and silhouette (Appendix A Table A2), showing the clusters are very cohesive and with a good to very good clustering quality. When we ran d-means for 150 users (Fig. 11(b) and Appendix A Table A5), 50 clusters were formed. Several clusters seem to overlap and some users that were alone in a cluster or with other users in Fig. 11(a), were assigned to different clusters, as more similar users appeared, rearranging the clusters and creating new ones. For instance, users 18 and 20 left C10 and were assigned to another cluster, C1, with higher similarities, and user 47 remained alone. As for k-means and k-means++ (Fig. 11(c)(d)(e)(f)), for both sets of users, we can see the members are more dispersed inside the same cluster, having several members far from the respective cluster’s centroid. We can also see there are more users closer to the respective cluster’s centroid for n=150 than for n=50, but there is also supposedly a cluster overlap. The members of each cluster are exactly the same for k-means and k-means++ for n=50. The same almost happens for n=150, but some changes occurred like users 62, 21, 144, 119, 52, and 113 changed from one cluster to the other, and the overlap seems to diminish, which may be the result of the k-means++ clusters initialization based on probabilities. Comparing the clustering metrics for the three algorithms (see Appendix A Table A2-A7), d-means clearly outperforms the baselines, showing high cohesive clusters and a good to very good clustering quality for both sets of users. In comparison, all the baselines, for both sets of users, have lower cohesion values and a weak clustering quality. This confirms d-means fulfilled its intended objective of clustering only the most similar users. 5.2. Simulation’s sample characterization As can be seen in Table 6, the simulation participants’gender is almost equilibrated (57% males vs 43% females) with a mean age of 34 years old, being most of them in the 20–35 (54%) age range (Fig. 12 (a)). 54% of the participants are in some sort of a relationship and the majority do not have children (63%). The great majority have higher education (86%) and are from the Engineering &Technology formation area, as most of the participants were from ISEP. Regarding the professional situation, the great majority (63%) have a job, 34% are only students, and 3% are unemployed. 26% are afraid of heights and 6% are claustrophobic. In Fig. 12(b), it can be seen that the participants who are claustrophobic are also afraid of heights. None of the respondents had the considered disabilities (physical impairments, sensorial (deafness, blindness), heart conditions), therefore, they were not an influencing factor for the group recommendations. The participants personality, assessed with the BFI, revealed the same type of distribution as in the previous studies (Alves, Martins, Saraiva, et al., 2023; Alves et al., 2020) (see Fig. 13), with the same 3 dimensions, openness, conscientiousness and agreeableness, having responses above the mid-point, revealing a negative skewness, i.e., participants situated themselves more between “3Neither agree nor disagree”and “5-Agree strongly”, reflecting the same social desirability bias common in self-reporting questionnaires (Pedregon et al., 2012), e. g. like the desire of being more intellectual in the case of openness to experience; effortful, self-effective and truthful in the case of conscientiousness; and moral and kind in the case of agreeableness. The other two dimensions, extraversion and neuroticism, had the mean value near the scale mid-point. All five distributions follow the shape of a normal curve, and according to the values of skewness and kurtosis obtained, and respective standard errors, although some having a slight skew and kurtosis, they are in acceptable ranges and the data are considered not Fig. 10. (a) Boxplots comparing the 3 algorithms in terms of the clusters’users’average distance to the respective centroid; (b) closest cluster; (c) silhouette values, from 35 to 100 k users, for sim ≥0.80; and (d) d-means number of users per cluster for each sensitivity, from 1000 to 100 k users. Only clusters with 2 or more users were considered. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 18 Fig. 11. Comparison of the clusters 2D spatial distribution for d-means and the 2 baselines, for n =50 and n =150. Fig. 12. Participants’(a) age range, (b) fears/phobias; n=35. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 19 significantly different from a normal distribution (Field, 2013; Gravetter, Wallnau, Forzano, &Witnauer, 2020; Sposito, Hand, &Skarpness, 1983). 5.3. Simulation results and analysis For the 35 participants, the dynamic clustering algorithm, d-means (Algorithm 1) created 21 clusters at the registration phase, as can be observed in Table 7. As expected in large groups, a great number of clusters was formed, revealing their members heterogeinity. However, even in a small sample, 23 of 35 participants (66%) could be fitted into the same cluster (9 clusters in total). This is curious, meaning 66% of the random participants had a personality similarity above 0.80 (Fig. 14(c)), which is reflected in the participants’personality scores, predicted factors’scores (Table 7), and in the radar charts of Fig. 17, for each cluster. This could be confirmed by the similarity measure between the participants inside each cluster. For example, sim(p8,p13)is 0.83, and sim(p8, p4)is 0.85 (cluster C11). This similarity can be due to most of the participants being from the same formation area. The participants who did not match the similarity criteria (Algorithm 1) were assigned to a new cluster or remained in their individual cluster if already in one. For example, users 28 and 34 were assigned to the same cluster C7but user 26 remained alone in a cluster, C4, because there were no participants with sim(p26,pj)≥0.80. With the d-means dynamic clustering process, there was no need to define the initial number of clusters required, as is mandatory for the existing clustering algorithms, since the number of clusters were dynamically generated/updated at the participants’registration process, according to the clustering iterations and the 0.80 similarity restriction. In Table 8 and Fig. 14, we can see the clusters have a low α value showing a good compactness and that they are very close to the respective centroid (Euclidean similarity close to 1). It can also be seen that the clusters that aggregated more participants have the highest average distance to the respective cluster’s centroid, being less compact than the ones with less participants. Pertaining to the clusters’separability from the neighbor cluster, there are closer clusters where the participants could have been assigned to. However, d-means guarantees the users are clustered with the most similar ones (it is possible, but improbable, to have several clusters with equal similarity as neighbor clusters, so the first one is chosen). The participants that got alone in a cluster are not shown (n=11). These participants affect the closest cluster and silhouette values, as they are considered clustering possibilities for the d-means algorithm, but they are not aggregated because they do not match the similarity criteria of at least 0.80. We also do not merge close clusters, to preserve the Fig. 13. Distribution of the participants’five personality dimensions scores. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 20 Fig. 14. (a) Participants’clusters’average α ;(b) Participants’clusters’silhouette; (c) Participants’Euclidean similarity to the corresponding cluster’s centroid, grouped by their respective cluster. Only clusters with more than one member are shown. Fig. 15. (a) Formed subgroups average distance to the corresponding subgroup’s centroid; (b) Formed subgroups silhouette values. Fig. 16. Example of some rules found with the Apriori algorithm. (a) 71% users who like to do hiking/mountaineering, visit beaches for their natural beauty and walk along the sea/river coast, like to appreciate natural landscapes, with a 100% confidence; (b) 71% users who like to walk along the sea/river coast and do not like to go to the circus, like to appreciate natural landscapes, with a 98% confidence; (c) 71% users who like to visit archaeological sites/ruins and taste typical dishes of the destination and visit monuments, like to visit historic cities/villages, with a 98% confidence. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 21 members’similarity as much as possible. Nevertheless, the formed clusters have very high silhouette values, all superior to 0.60 and most of them with s≥0.80 (Fig. 14(b)), showing a great clustering quality. We are not forcing users that are alone to cluster with other users like usually happens in the other clustering algorithms, to have higher cohesions, which is the purpose of d-means, to have the most similar personality users in the same cluster. The fact we force a minimum similarity greatly improves the tourists’proximity in the same cluster (Fig. 14(c)), guarantying that tourists with very different personalities are not grouped together like may happen in k-means and k-means++, eliminating the noise and outlier issues associated to the baseline algorithms. When the participants asked for group recommendations, the MAMS detected that the excursion group could be divided into different subgroups, forcing the main group members that were alone or with only another member from the same cluster to pair with each other or other subgroups, being assigned to the one with the closest centroid until forming subgroups of at least 3 members. Seven subgroups were created, as can be seen in Fig. 6(b) and Table 9, meaning the participants were sufficiently different to divide the main group. Also, at least 2 participants from the same cluster were placed in the same subgroup, meaning they had similar preferences for the 0.80 similarity threshold (Table 7). For example, participant p1(from cluster C21) was forced to pair with participants p4,p8, and p13 (from C11) in subgroup C11Sg, as he was more similar to their subgroup’s centroid than to the others. This led to some subgroups ending up with tourists not so similar to the other members, but that were the most similar possible, like subgroup C7Sg. As there was not a group member that could be grouped with participants p7, and p23, that came from the same cluster C17, they were separated into different subgroups, C3Sg and C14Sg respectively, being assigned to the subgroup with the most similar centroid personality, as imposed by the implemented algorithm. To determine the clustering quality of the subgroups formed, their compactness, separability, and silhouette values were calculated, revealing very compact subgroups with a very good clustering quality, all with s>0.80 with an average silhouette s=0.91 (see Fig. 15 and Appendix A Table A1), confirming the intra-subgroups personality similarity. The formation of the subgroups revealed to be advantageous as the subgroups (clusters) formed became more compact and with higher silhouettes than the originating clusters. The formed subgroups can be better visualized in the radar charts at Fig. 17. For each subgroup, the similarity between the participants belonging to the same cluster can Fig. 17. Subgroups formed displayed as a radar chart of the corresponding members’personality. The participants that came from the same cluster are in bold. See Table 7 for detailed information on the subgroups. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 22 be perfectly seen. Also, the personality of the participants that came from another cluster(s) is close to their subgroup’s members, which is in line with the obtained silhouette values. As for the rules, Fig. 16 shows some of the rules found with the Apriori algorithm run by the MAMS for the 1035 users from the previous study dataset (Alves, Martins, Saraiva, et al., 2023). In 10 000 rules, most were between the tourist attractions preference. No rules relating to the other attributes, such as age, marital status, children, etc., could be found in the rules set. This may be due to the large number of variables in comparison, or some variables needed to be differently binned, being a subject for further research. All the rules found had a great confidence level, with 119 rules having a 100% confidence, 568 rules 99%, and so on, until a 91% confidence. To validate the algorithms correctness, the Apriori algorithm was also run in Weka®3.8.6 for the same sample, resulting in the same rules. The generated rules can be consulted at https://www.gecad.isep.ipp.pt/grouplanner/dissemina tion.html under the “simulations”tab. The average personality and predicted tourism categories preference for the subgroups formed can be consulted in Table 10 (Left side). 8 Table 10 right side shows the average ratings participants gave to the same categories in the pre-questionnaire. The list of POI suggested for each subgroup and the respective average rating can be consulted in Table 9. The individual ratings given by each group member to each recommended POI can be consulted in Appendix B Table B2. The predicted tourism categories that subgroup C11Sg scored higher and with a score of at least 0.49 were only two, “Cultural Heritage”and “Museums, Boat Trips &Viewpoints”. The 3 predicted categories C3Sg and C9Sg scored higher were “Cultural Heritage”,“Museums, Boat Trips &Viewpoints”and “Wild Nature”, respectively. C7Sg had higher predicted values for “Cultural Heritage”,“Wild Nature”and “Museums, Boat Trips &Viewpoints”. For C14Sg, the predicted top-3 categories were “Cultural Heritage”,“Wild Nature”and “Natural Phenomena”. Finally, C15Sg and C5Sg had a higher predicted score on “Cultural Heritage”, “Museums, Boat Trips &Viewpoints”and “Natural Phenomena”, being those categories the ones selected for the preliminary recommendations. Although subgroups C3Sg and C9Sg, and C15Sg and C5Sg, have the same top-3 categories, they have different personalities, which is the reason why they are distinct subgroups, as we believe it is important to preserve the intra-subgroup’s similarity to mitigate the heterogeneity problems associated with large excursion groups, as the members will have more similar preferences and therefore can reach a faster consensus in the final recommendation list. To visit monuments and historical cities/villages (F7) are the predicted top preferred activities for all subgroups. To visit museums, go on a boat trip or visit viewpoints (F5), activities in the wild nature (F2) or to appreciate natural phenomena (F11), are the next predicted most preferred tourist attractions. These preferences were also the participants’top-rated tourism categories in the pre-questionnaire. 9 This is in line with the results previously found (Alves, Martins, Saraiva, et al., 2023), that, regardless the person’s personality, everyone likes to visit/ do those types of tourist attractions when on vacation, along with theme &animal parks (F6), relax/swim at the beach/swimming pool (F4), taste typical dishes (F9), and party, music &nightlife events (F3). Only subgroups C7Sg and C3Sg would like to experience adrenaline activities (F1), approximately to the predicted. Sports &games (F8) had a negative prediction for all subgroups, although C14Sg and C5Sg would like to include that sort of activities in their vacations as mentioned in the prequestionnaire. Nonetheless, it is easy to observe that the predicted scores for other categories in a subgroup only differ one-tenth from the ones suggested for the top. For instance, C7Sg could easily have F9,F11,F1or even F3and F4as top categories, confirming the given ratings in the prequestionnaire. The same happens for C3Sg,C14Sg and C9Sg for the same or other categories. This means a small variation in the BFI answers can change that order, and that the suggested POI can be further refined by using the ratings given by the users when using the application. This proximity can also be used to change the recommended POI variability. To validate these results and determine if the predicted ratings were not statistically different from the real ratings given in the prequestionnaire, the paired samples t-test to compare their means was performed in SPSS. The results can be found in Table 11. The predicted ratings for the tourism categories F1(Adrenaline Activities), F5(Museums, Boat trips &Viewpoints), F7(Cultural heritage), F8(Sports & Games), F10 (Health &Well-being) and F11 (Natural phenomena) were not statistically different from the real ratings. The same applies to the travel-related concerns FP1(Previsibility &Safety) and FP4(Familiarity). For the other factors, the difference is small for F3(Party, Music & Nightlife), and medium for F9(Gastronomy events) and FP3(Uniqueness &Exoticness) (Cohen, 2013), and therefore can be considered for this study and context, meaning 11 of the 15 categories had valid predictions, validating the used structural equation models (Alves, Martins, Saraiva, et al., 2023). Comparing the given to the predicted preferences for the tourism categories (Table 10 left and right sides), in a 1 to 5 stars scale, almost all categories (F1,F3,F5,F7,F8,F9,F10 and F11) have similar mean ratings, confirming the predictions. F4and F6were the less accurate predicted categories, followed by F2, which is reflected in the paired samples t-test results (Table 11), having a high effect size. We can also see that no tourism category had an average predicted preference or given rating lower than 2. This may be a consequence of the small sample used for the simulation, as it was impossible to have a significant number of participants for each type of personality combination. For example, if there were a greater number of participants with low openness, conscientiousness or agreeableness values, or higher neuroticism, the categories that were less chosen would increase their scores for those participants, namely F1,F3,F8, or F10. The same applies to the predicted ones. Table 8 Clusters formed by the proposed d-means algorithm and their respective average distance to the centroid ( α ), closest cluster (β, separability), silhouette values (s), and Euclidean similarity (sim) between the participant piand the corresponding cluster’s centroid. Only clusters with more than one member are shown. piclusterCq.id α Closest cluster, β sim(pi,Cq.centroid)Silhouette,s 5C30,0189 0,0511 0,927678125 0,6303 6C30,0189 0,0511 0,87106191 0,6303 25 C30,0189 0,0511 0,855179517 0,6303 11 C90,0183 0,0520 0,945643965 0,6491 15 C90,0183 0,0520 0,889973882 0,6491 32 C90,0183 0,0520 0,889399342 0,6491 19 C90,0183 0,0520 0,868055569 0,6491 33 C90,0183 0,0520 0,844076439 0,6491 7C17 0,0135 0,0693 0,895905223 0,8052 23 C17 0,0135 0,0693 0,895905223 0,8052 8C11 0,0135 0,0838 0,909922294 0,8385 4C11 0,0135 0,0838 0,894875885 0,8385 13 C11 0,0135 0,0838 0,884655094 0,8385 35 C14 0,0125 0,0786 0,899530081 0,8413 31 C14 0,0125 0,0786 0,899530081 0,8413 17 C15 0,0075 0,0511 0,920422077 0,8537 12 C15 0,0075 0,0511 0,920422077 0,8537 28 C70,0099 0,0786 0,909609356 0,8744 34 C70,0099 0,0786 0,909609356 0,8744 30 C50,0066 0,0712 0,925127505 0,9081 16 C50,0066 0,0712 0,925127505 0,9081 22 C60,0041 0,0520 0,939994755 0,9217 14 C60,0041 0,0520 0,939994755 0,9217 8 More details on how the predicted scores were calculated can be consulted in Alves, Martins, Saraiva, et al. (2023),Section 5. 9 The individual ratings given by each participant to the tourism categories and travel-related preferences &concerns can be consulted in Appendix B Table B. 1. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 23 Analyzing Table 10 right side, all subgroups want to have cultural & learning experiences (FP2) and visit unique &exotic places (FP3), especially C7Sg,C3Sg and C15Sg. It was only for one-tenth that all those preferences would have been predicted for the suggested POI, as can be observed on the left side of the table, also confirming the paired samples t-test results. All subgroups want safety &previsibility (FP1) when on vacations, except C7Sg and C14Sg which were neutral. Only C9Sg and C5Sg are concerned about familiarity (FP4) and would prefer not to travel with strangers. The categories with mean differences not statistically different at a significance level of α =0.025 (2-tailed) are in bold. The Cohen’s small effect sizes are in bold, the medium are underlined, and the remaining are the high effect sizes (Cohen, 2013). As most of the members of C15Sg were afraid of heights (in fact all members), the REMS did not suggest POI for acrophobic, and therefore only 2 were available in the POIMS under those criteria and thus recommended (Table 9). Some participants referred they did not agree with the predicted tourism categories individual preference order (screen in Fig. 2(d) from the user’s profile in the app). This reflects the same observations already made for the subgroups’predicted preferences. Besides, these are only the initial recommendations which will be improved with the rules, the ratings given by the users to the recommendation lists and already visited POI. Also, all the presented predictions are a reflection of the models used (Alves, Martins, Saraiva, et al., 2023), which had an acceptable fit (CFI =0.771 and PCFI =0.721 for the “Personality vs Tourism Categories Preference model”, and CFI =0.809 and PCFI = 0.716 for the “Personality vs Travel-related Preferences &Concerns” model, among other metrics). A larger sample (at least the triple (Marˆ oco, 2010)) was needed to increase the models’goodness-of-fit and Table 9 List of POI recommended to each subgroup with the respective POI tourism category, POI Id, POI description and average rating given by the simulation participants (from 1 to 5 stars). Subgroup 11 Subgroup 3 Category Id Name Mean rating Category Id Name Mean rating F7 652 Pal´ acio da Bolsa 4F7 652 Pal´ acio da Bolsa 4 F7 885 S´ e Catedral do Porto 4F7 885 S´ e Catedral do Porto 4 F7 382 Igreja de S˜ ao Francisco 3 F7 382 Igreja de S˜ ao Francisco 3 F7 170 Centro Hist´ orico de Guimar˜ aes 4F7 170 Centro Hist´ orico de Guimar˜ aes 3 F7 135 Castelo de Guimar˜ aes 3 F5 588 Museu FC Porto 4 F5 606 Museu Serralves 5F5 624 Navio Gil Eannes 4 F5 588 Museu FC Porto 3 F5 174 Centro Portuguˆ es de Fotografia 4 F5 624 Navio Gil Eannes 3 F2 486 Mata Nacional do Buçaco 4 F5 174 Centro Portuguˆ es de Fotografia 4F2 519 MontanhaParque da Penha 4 F5 580 Museu do Douro 4F2 33 Arouca Geopark 3 Subgroup 7 Subgroup 14 Category Id Name Mean rating Category Id Name Mean rating F7 57 Bom Jesus do Monte 5F7 652 Pal´ acio da Bolsa 5 F7 652 Pal´ acio da Bolsa 5F7 885 S´ e Catedral do Porto 4 F7 885 S´ e Catedral do Porto 4F7 382 Igreja de S˜ ao Francisco 3 F7 382 Igreja de S˜ ao Francisco 5F7 170 Centro Hist´ orico de Guimar˜ aes 5 F2 486 Mata Nacional do Buçaco 4F2 683 Passadiços do Paiva 5 F2 519 MontanhaParque da Penha 4F2 486 Mata Nacional do Buçaco 5 F2 283 Excurs˜ ao de meio dia ao Pal´ acio Bussaco, jardins e floresta 5F2 33 Arouca Geopark 5 F5 588 Museu FC Porto 5F11 690 Pateira de Fermentelos 5 F5 624 Navio Gil Eannes 4F11 708 Poço Negro 5 F5 174 Centro Portuguˆ es de Fotografia 3 F11 692 Pedra Bolideira 2 Subgroup 9 Subgroup 5 Category Id Name Mean rating Category Id Name Mean rating F7 652 Pal´ acio da Bolsa 4F7 652 Pal´ acio da Bolsa 4 F7 885 S´ e Catedral do Porto 4F7 885 S´ e Catedral do Porto 4 F7 382 Igreja de S˜ ao Francisco 3 F7 382 Igreja de S˜ ao Francisco 4 F7 170 Centro Hist´ orico de Guimar˜ aes 5F7 170 Centro Hist´ orico de Guimar˜ aes 5 F5 588 Museu FC Porto 2 F5 588 Museu FC Porto 3 F5 624 Navio Gil Eannes 4F5 624 Navio Gil Eannes 4 F5 174 Centro Portuguˆ es de Fotografia 3 F5 174 Centro Portuguˆ es de Fotografia 4 F2 683 Passadiços do Paiva 4F11 690 Pateira de Fermentelos 5 F2 486 Mata Nacional do Buçaco 4F11 708 Poço Negro 5 F2 33 Arouca Geopark 4F11 692 Pedra Bolideira 3 Subgroup 15 Category Id Name Mean rating F7 885 S´ e Catedral do Porto 4 F7 382 Igreja de S˜ ao Francisco 4 F7 170 Centro Hist´ orico de Guimar˜ aes 4 F7 135 Castelo de Guimar˜ aes 4 F5 588 Museu FC Porto 2 F5 624 Navio Gil Eannes 3 F5 174 Centro Portuguˆ es de Fotografia 4 F11 690 Pateira de Fermentelos 5 F11 457 Lagoas Naturais Fafi˜ ao 5 Values in bold represent positively rated POI and bold underlined POI with a rating of 5 stars. The ratings in a subgroup correspond to the average rating given by the members belonging only to the same cluster. POI names are in Portuguese. The individual ratings given by each subgroup member can be consulted in Appendix A Table B2. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 24 consequently the predictions. However, we believe it is a great help and contribution to GRS for tourism, and in general, to be able to present these first recommendations to excursion groups, as well as divide them into subgroups of similar personalities, mitigating the cold-start problem and diminishing the excursion group’s heterogeneity and conflicting preferences. This is confirmed by the high ratings the participants gave to the individual POI recommendation list (Fig. 18(a)): 74% of the recommended POI had 4 to 5 stars, and if we include ratings from 3 to 5 stars, 92%, meaning the participants were highly satisfied with the recommendations suggested. Only 8% of the POI were not an option to visit. It is important to mention that during the simulation some participants gave lower rates to POI they already visited (because they didn’t like or just didn’t want to revisit) or just because they didn’t know what the POI was. These situations can be easily resolved if we add the option to indicate if the POI has already been visited and a representative picture of the POI to the recommendation lists (under development). According to the MAMS implemented algorithms, if the group owner is asked to retrieve new recommendations, all the POI with a rating below 3 (configurable) will be replaced by new ones. According to this criterion, only subgroups C14Sg,C9Sg and C15Sg would have POI replaced, namely the “588-FC Porto Museum”for subgroups C9Sg and C15Sg, and “692-Boulder Stone”for subgroup C14Sg. This is a great result, meaning 96% of the participants were satisfied with the suggested POI group recommendations (Fig. 18(b)). If we also replaced the POI rated with 3 stars, 75% of the participants would be very satisfied. Interestingly, the preference for visiting the FC Porto Museum was below 3 stars for several participants, either from different clusters or Table 10 Left side: predicted average preference for the tourism categories and the travel-related preferences &concerns for the formed subgroups (CqSg), from 0.0 to 1.0, followed by the corresponding rating from 1 to 5 stars in parenthesis. Right side: average ratings for the tourism categories and travel-related preferences &concerns given by the members in the pre-questionnaire, from 1 to 5 stars. Both sides show the subgroups average personality for easier comparison, from 0.0 to 1.0 (only calculated for members belonging to the same cluster). Predicted average preferences Average ratings given C11Sg C7Sg C3Sg C14Sg C9Sg C15Sg C5Sg C11Sg C7Sg C3Sg C14Sg C9Sg C15Sg C5Sg O0.77 0.86 0.73 0.69 0.61 0.70 0.55 O0.77 0.86 0.73 0.69 0.61 0.70 0.55 C0.72 0.70 0.64 0.63 0.84 0.52 0.50 C0.72 0.70 0.64 0.63 0.84 0.52 0.50 E0.29 0.83 0.59 0.77 0.50 0.42 0.30 E0.29 0.83 0.59 0.77 0.50 0.42 0.30 A0.51 0.82 0.65 0.84 0.65 0.59 0.71 A0.51 0.82 0.65 0.84 0.65 0.59 0.71 N0.56 0.60 0.53 0.39 0.42 0.68 0.47 N0.56 0.60 0.53 0.39 0.42 0.68 0.47 F1 0.29 (2) 0.67 (4) 0.52 (3) 0.64 (4) 0.39 (3) 0.44 (3) 0.36 (3) F1 34 4 3 2 2 2 F2 0.39 (2) 0.73 (4) 0.58 (3) 0.73 (4) 0.51 (4) 0.51 (3) 0.53 (3) F2 4 5 4 4 4 4 4 F3 0.32 (2) 0.66 (4) 0.52 (3) 0.62 (4) 0.44 (3) 0.44 (3) 0.35 (2) F3 4 5 4 4 3 3 3 F4 0.30 (2) 0.64 (4) 0.50 (3) 0.62 (4) 0.46 (3) 0.41 (3) 0.35 (2) F4 5 5 5 4 4 5 4 F5 0.49 (3) 0.69 (4) 0.59 (3) 0.68 (4) 0.52 (3) 0.58 (3) 0.63 (4) F5 4 4 4 4 4 4 4 F6 0.33 (2) 0.61 (4) 0.50 (3) 0.58 (3) 0.44 (3) 0.43 (3) 0.35 (2) F6 4 5 4 4 4 4 4 F7 0.51 (3) 0.82 (5) 0.65 (4) 0.84 (5) 0.65 (4) 0.59 (3) 0.71 (4) F7 4 4 4 4 5 4 5 F8 0.33 (2) 0.54 (3) 0.47 (3) 0.53 (3) 0.41 (3) 0.41 (3) 0.34 (2) F8 3 2 3 43 3 4 F9 0.33 (2) 0.68 (4) 0.53 (3) 0.68 (4) 0.48 (3) 0.45 (3) 0.43 (3) F9 4 4 4 4 4 34 F10 0.37 (2) 0.62 (4) 0.51 (3) 0.57 (3) 0.48 (3) 0.44 (3) 0.35 (2) F10 4 4 3 3 43 2 F11 0.39 (2) 0.68 (4) 0.55 (3) 0.69 (4) 0.47 (3) 0.51 (3) 0.54 (4) F11 4 4 4 4 4 5 3 FP1 0.55 (3) 0.80 (4) 0.64 (4) 0.72 (4) 0.58 (3) 0.60 (3) 0.57 (3) FP1 4 3434 4 4 FP2 0.57 (3) 0.55 (3) 0.53 (3) 0.47 (3) 0.58 (3) 0.54 (3) 0.48 (3) FP2 4 4 4 4 4 4 4 FP3 0.48 (3) 0.83 (5) 0.64 (4) 0.77 (4) 0.57 (3) 0.54 (3) 0.49 (3) FP3 4 5 5 4 4 5 4 FP4 0.59 (3) 0.43 (3) 0.48 (3) 0.36 (2) 0.53 (3) 0.50 (3) 0.40 (2) FP4 2 3 3 3 434 Fears heights Left side: The values in bold underlined represent the top tourism categories (F1to F11) for each subgroup superior or equal to 0.49. The values in bold represent the travel-related preferences &concerns (FP1to FP4) superior or equal to 0.50. The scale binning corresponds to: 0.0–0.20 (1), 0.21–0.40(2), 0.41–0.60(3), 0.61–0.80(4), 0.81–1.0(5). Right side: The values in bold represent positively rated factors and bold underlined the ones rated with 5. The presented fears are only for subgroups where at least 50% of the members said to have that fear. Table 11 Paired samples t-test of the means between the real ratings given by the participants to the tourism categories and the travel-related preferences &concerns in the prequestionnaire and the predicted ones. 95% Confidence Interval of the Difference Pairs Mean Std. Deviation Std. Error Mean Lower Upper t df Sig. (2-tailed) Cohen’s D F1real −F1predicted ¡0,057 1,282 0,217 ¡0,498 0,383 ¡0,264 34 0,794 ¡0,04 F2real −F2predicted 0,857 1,141 0,193 0,465 1,249 4,443 34 0,000 0,75 F3real −F3predicted 0,514 1,040 0,176 0,157 0,871 2,927 34 0,006 0,49 F4real −F4predicted 1,486 1,197 0,202 1,074 1,897 7,341 34 0,000 1,24 F5real −F5predicted 0,286 0,926 0,156 ¡0,032 0,604 1,826 34 0,077 0,31 F6real −F6predicted 0,943 0,906 0,153 0,632 1,254 6,159 34 0,000 1,04 F7real −F7predicted 0,314 0,796 0,135 0,041 0,588 2,336 34 0,026 0,39 F8real −F8predicted 0,343 1,552 0,262 ¡0,190 0,876 1,307 34 0,200 0,22 F9real −F9predicted 0,543 1,094 0,185 0,167 0,919 2,936 34 0,006 0,50 F10 real −F10 predicted 0,400 1,193 0,202 ¡0,010 0,810 1,983 34 0,055 0,34 F11 real −F11 predicted 0,371 1,262 0,213 ¡0,062 0,805 1,741 34 0,091 0,29 FP1real −FP1predicted 0,171 0,985 0,166 ¡0,167 0,510 1,030 34 0,310 0,17 FP2real −FP2predicted 1,086 0,951 0,161 0,759 1,412 6,755 34 0,000 1,14 FP3real −FP3predicted 0,629 0,910 0,154 0,316 0,941 4,085 34 0,000 0,69 FP4real −FP4predicted 0,400 1,311 0,222 ¡0,050 0,850 1,806 34 0,080 0,31 P. Alves et al. Expert Systems With Applications 255 (2024) 124812 25 Table B2 (continued) User # Subgroup POI ID POI name Rating/Interest 35 C14Sg 33 Arouca Geopark 4 35 C14Sg 690 Pateira de Fermentelos 4 35 C14Sg 708 Poço Negro 4 35 C14Sg 692 Pedra Bolideira 1 10 C15Sg 588 Museu FC Porto 1 10 C15Sg 624 O Navio Gil Eannes 3 10 C15Sg 174 Centro Portuguˆ es de Fotografia 2 10 C15Sg 580 Museu do Douro 3 10 C15Sg 885 S´ e Catedral do Porto 3 10 C15Sg 382 Igreja de S˜ ao Francisco 3 10 C15Sg 170 Centro Hist´ orico de Guimar˜ aes 4 10 C15Sg 690 Pateira de Fermentelos 5 10 C15Sg 457 Lagoas Naturais Fafi˜ ao 5 36 C14Sg 652 Pal´ acio da Bolsa 5 36 C14Sg 885 S´ e Catedral do Porto 4 36 C14Sg 382 Igreja de S˜ ao Francisco 3 36 C14Sg 170 Centro Hist´ orico de Guimar˜ aes 5 36 C14Sg 683 Passadiços do Paiva 5 36 C14Sg 486 Mata Nacional do Buçaco 3 36 C14Sg 33 Arouca Geopark 5 36 C14Sg 690 Pateira de Fermentelos 3 36 C14Sg 708 Poço Negro 2 36 C14Sg 692 Pedra Bolideira 3 32 C9Sg 652 Pal´ acio da Bolsa 4 32 C9Sg 885 S´ e Catedral do Porto 3 32 C9Sg 382 Igreja de S˜ ao Francisco 2 32 C9Sg 170 Centro Hist´ orico de Guimar˜ aes 4 32 C9Sg 588 Museu FC Porto 1 32 C9Sg 624 O Navio Gil Eannes 4 32 C9Sg 174 Centro Portuguˆ es de Fotografia 3 32 C9Sg 683 Passadiços do Paiva 4 32 C9Sg 486 Mata Nacional do Buçaco 4 32 C9Sg 33 Arouca Geopark 4 23 C14Sg 652 Pal´ acio da Bolsa 4 23 C14Sg 885 S´ e Catedral do Porto 2 23 C14Sg 382 Igreja de S˜ ao Francisco 3 23 C14Sg 170 Centro Hist´ orico de Guimar˜ aes 5 23 C14Sg 683 Passadiços do Paiva 5 23 C14Sg 486 Mata Nacional do Buçaco 5 23 C14Sg 33 Arouca Geopark 5 23 C14Sg 690 Pateira de Fermentelos 4 23 C14Sg 708 Poço Negro 5 23 C14Sg 692 Pedra Bolideira 3 22 C3Sg 652 Pal´ acio da Bolsa 5 22 C3Sg 885 S´ e Catedral do Porto 5 22 C3Sg 382 Igreja de S˜ ao Francisco 4 22 C3Sg 170 Centro Hist´ orico de Guimar˜ aes 4 22 C3Sg 588 Museu FC Porto 3 22 C3Sg 624 O Navio Gil Eannes 4 22 C3Sg 174 Centro Portuguˆ es de Fotografia 4 22 C3Sg 486 Mata Nacional do Buçaco 4 22 C3Sg 519 MontanhaParque da Penha 4 22 C3Sg 33 Arouca Geopark 3 21 C5Sg 652 Pal´ acio da Bolsa 5 21 C5Sg 885 S´ e Catedral do Porto 5 21 C5Sg 382 Igreja de S˜ ao Francisco 5 21 C5Sg 170 Centro Hist´ orico de Guimar˜ aes 4 21 C5Sg 588 Museu FC Porto 5 21 C5Sg 624 O Navio Gil Eannes 4 21 C5Sg 174 Centro Portuguˆ es de Fotografia 2 21 C5Sg 690 Pateira de Fermentelos 4 21 C5Sg 708 Poço Negro 5 21 C5Sg 692 Pedra Bolideira 3 1C11Sg 652 Pal´ acio da Bolsa 3 1C11Sg 885 S´ e Catedral do Porto 4 1C11Sg 382 Igreja de S˜ ao Francisco 5 1C11Sg 170 Centro Hist´ orico de Guimar˜ aes 4 1C11Sg 135 Castelo de Guimar˜ aes 5 1C11Sg 606 Museu Serralves 3 1C11Sg 588 Museu FC Porto 2 1C11Sg 624 O Navio Gil Eannes 2 1C11Sg 174 Centro Portuguˆ es de Fotografia 3 1C11Sg 580 Museu do Douro 2 13 C11Sg 652 Pal´ acio da Bolsa 4 13 C11Sg 885 S´ e Catedral do Porto 4 13 C11Sg 382 Igreja de S˜ ao Francisco 5 (continued on next page) P. Alves et al. Expert Systems With Applications 255 (2024) 124812 32 Table B2 (continued) User # Subgroup POI ID POI name Rating/Interest 13 C11Sg 170 Centro Hist´ orico de Guimar˜ aes 5 13 C11Sg 135 Castelo de Guimar˜ aes 4 13 C11Sg 606 Museu Serralves 5 13 C11Sg 588 Museu FC Porto 2 13 C11Sg 624 O Navio Gil Eannes 3 13 C11Sg 174 Centro Portuguˆ es de Fotografia 5 13 C11Sg 580 Museu do Douro 4 15 C9Sg 652 Pal´ acio da Bolsa 4 15 C9Sg 885 S´ e Catedral do Porto 3 15 C9Sg 382 Igreja de S˜ ao Francisco 2 15 C9Sg 170 Centro Hist´ orico de Guimar˜ aes 5 15 C9Sg 588 Museu FC Porto 4 15 C9Sg 624 O Navio Gil Eannes 5 15 C9Sg 174 Centro Portuguˆ es de Fotografia 3 15 C9Sg 683 Passadiços do Paiva 5 15 C9Sg 486 Mata Nacional do Buçaco 3 15 C9Sg 33 Arouca Geopark 4 31 C14Sg 652 Pal´ acio da Bolsa 5 31 C14Sg 885 S´ e Catedral do Porto 4 31 C14Sg 382 Igreja de S˜ ao Francisco 3 31 C14Sg 170 Centro Hist´ orico de Guimar˜ aes 4 31 C14Sg 683 Passadiços do Paiva 5 31 C14Sg 486 Mata Nacional do Buçaco 5 31 C14Sg 33 Arouca Geopark 5 31 C14Sg 690 Pateira de Fermentelos 5 31 C14Sg 708 Poço Negro 5 31 C14Sg 692 Pedra Bolideira 3 9C14Sg 652 Pal´ acio da Bolsa 3 9C14Sg 885 S´ e Catedral do Porto 2 9C14Sg 382 Igreja de S˜ ao Francisco 2 9C14Sg 170 Centro Hist´ orico de Guimar˜ aes 4 9C14Sg 683 Passadiços do Paiva 4 9C14Sg 486 Mata Nacional do Buçaco 4 9C14Sg 33 Arouca Geopark 4 9C14Sg 690 Pateira de Fermentelos 3 9C14Sg 708 Poço Negro 4 9C14Sg 692 Pedra Bolideira 3 33 C9Sg 652 Pal´ acio da Bolsa 5 33 C9Sg 885 S´ e Catedral do Porto 3 33 C9Sg 382 Igreja de S˜ ao Francisco 3 33 C9Sg 170 Centro Hist´ orico de Guimar˜ aes 5 33 C9Sg 588 Museu FC Porto 2 33 C9Sg 624 O Navio Gil Eannes 4 33 C9Sg 174 Centro Portuguˆ es de Fotografia 3 33 C9Sg 683 Passadiços do Paiva 1 33 C9Sg 486 Mata Nacional do Buçaco 3 33 C9Sg 33 Arouca Geopark 3 28 C7Sg 57 Bom Jesus do Monte 5 28 C7Sg 652 Pal´ acio da Bolsa 5 28 C7Sg 885 S´ e Catedral do Porto 4 28 C7Sg 382 Igreja de S˜ ao Francisco 5 28 C7Sg 486 Mata Nacional do Buçaco 3 28 C7Sg 519 MontanhaParque da Penha 3 28 C7Sg 283 Excurs˜ ao de meio dia ao Pal´ acio Bussaco, jardins e floresta 4 28 C7Sg 588 Museu FC Porto 5 28 C7Sg 624 O Navio Gil Eannes 4 28 C7Sg 174 Centro Portuguˆ es de Fotografia 3 18 C5Sg 652 Pal´ acio da Bolsa 4 18 C5Sg 885 S´ e Catedral do Porto 5 18 C5Sg 382 Igreja de S˜ ao Francisco 4 18 C5Sg 170 Centro Hist´ orico de Guimar˜ aes 5 18 C5Sg 588 Museu FC Porto 2 18 C5Sg 624 O Navio Gil Eannes 5 18 C5Sg 174 Centro Portuguˆ es de Fotografia 4 18 C5Sg 690 Pateira de Fermentelos 5 18 C5Sg 708 Poço Negro 5 18 C5Sg 692 Pedra Bolideira 4 17 C15Sg 588 Museu FC Porto 1 17 C15Sg 624 O Navio Gil Eannes 4 17 C15Sg 174 Centro Portuguˆ es de Fotografia 4 17 C15Sg 580 Museu do Douro 4 17 C15Sg 885 S´ e Catedral do Porto 4 17 C15Sg 382 Igreja de S˜ ao Francisco 4 17 C15Sg 170 Centro Hist´ orico de Guimar˜ aes 2 17 C15Sg 690 Pateira de Fermentelos 4 17 C15Sg 457 Lagoas Naturais Fafi˜ ao 4 (continued on next page) P. Alves et al. Expert Systems With Applications 255 (2024) 124812 33 Table B2 (continued) User # Subgroup POI ID POI name Rating/Interest 34 C7Sg 57 Bom Jesus do Monte 5 34 C7Sg 652 Pal´ acio da Bolsa 5 34 C7Sg 885 S´ e Catedral do Porto 4 34 C7Sg 382 Igreja de S˜ ao Francisco 5 34 C7Sg 486 Mata Nacional do Buçaco 5 34 C7Sg 519 MontanhaParque da Penha 5 34 C7Sg 283 Excurs˜ ao de meio dia ao Pal´ acio Bussaco, jardins e floresta 5 34 C7Sg 588 Museu FC Porto 4 34 C7Sg 624 O Navio Gil Eannes 4 34 C7Sg 174 Centro Portuguˆ es de Fotografia 3 20 C9Sg 652 Pal´ acio da Bolsa 5 20 C9Sg 885 S´ e Catedral do Porto 5 20 C9Sg 382 Igreja de S˜ ao Francisco 4 20 C9Sg 170 Centro Hist´ orico de Guimar˜ aes 5 20 C9Sg 588 Museu FC Porto 3 20 C9Sg 624 O Navio Gil Eannes 5 20 C9Sg 174 Centro Portuguˆ es de Fotografia 5 20 C9Sg 683 Passadiços do Paiva 5 20 C9Sg 486 Mata Nacional do Buçaco 5 20 C9Sg 33 Arouca Geopark 5 2C14Sg 652 Pal´ acio da Bolsa 4 2C14Sg 885 S´ e Catedral do Porto 4 2C14Sg 382 Igreja de S˜ ao Francisco 5 2C14Sg 170 Centro Hist´ orico de Guimar˜ aes 3 2C14Sg 683 Passadiços do Paiva 5 2C14Sg 486 Mata Nacional do Buçaco 5 2C14Sg 33 Arouca Geopark 5 2C14Sg 690 Pateira de Fermentelos 4 2C14Sg 708 Poço Negro 3 2C14Sg 692 Pedra Bolideira 3 25 C3Sg 652 Pal´ acio da Bolsa 5 25 C3Sg 885 S´ e Catedral do Porto 4 25 C3Sg 382 Igreja de S˜ ao Francisco 3 25 C3Sg 170 Centro Hist´ orico de Guimar˜ aes 5 25 C3Sg 588 Museu FC Porto 3 25 C3Sg 624 O Navio Gil Eannes 5 25 C3Sg 174 Centro Portuguˆ es de Fotografia 5 25 C3Sg 486 Mata Nacional do Buçaco 4 25 C3Sg 519 MontanhaParque da Penha 5 25 C3Sg 33 Arouca Geopark 5 16 C5Sg 652 Pal´ acio da Bolsa 3 16 C5Sg 885 S´ e Catedral do Porto 2 16 C5Sg 382 Igreja de S˜ ao Francisco 3 16 C5Sg 170 Centro Hist´ orico de Guimar˜ aes 5 16 C5Sg 588 Museu FC Porto 4 16 C5Sg 624 O Navio Gil Eannes 4 16 C5Sg 174 Centro Portuguˆ es de Fotografia 4 16 C5Sg 690 Pateira de Fermentelos 5 16 C5Sg 708 Poço Negro 5 16 C5Sg 692 Pedra Bolideira 3 12 C15Sg 588 Museu FC Porto 3 12 C15Sg 624 O Navio Gil Eannes 2 12 C15Sg 174 Centro Portuguˆ es de Fotografia 4 12 C15Sg 580 Museu do Douro 4 12 C15Sg 885 S´ e Catedral do Porto 4 12 C15Sg 382 Igreja de S˜ ao Francisco 3 12 C15Sg 170 Centro Hist´ orico de Guimar˜ aes 5 12 C15Sg 690 Pateira de Fermentelos 5 12 C15Sg 457 Lagoas Naturais Fafi˜ ao 5 14 C3Sg 652 Pal´ acio da Bolsa 4 14 C3Sg 885 S´ e Catedral do Porto 4 14 C3Sg 382 Igreja de S˜ ao Francisco 3 14 C3Sg 170 Centro Hist´ orico de Guimar˜ aes 3 14 C3Sg 588 Museu FC Porto 4 14 C3Sg 624 O Navio Gil Eannes 4 14 C3Sg 174 Centro Portuguˆ es de Fotografia 4 14 C3Sg 486 Mata Nacional do Buçaco 4 14 C3Sg 519 MontanhaParque da Penha 4 14 C3Sg 33 Arouca Geopark 3 19 C9Sg 652 Pal´ acio da Bolsa 3 19 C9Sg 885 S´ e Catedral do Porto 4 19 C9Sg 382 Igreja de S˜ ao Francisco 5 19 C9Sg 170 Centro Hist´ orico de Guimar˜ aes 4 19 C9Sg 588 Museu FC Porto 2 19 C9Sg 624 O Navio Gil Eannes 4 19 C9Sg 174 Centro Portuguˆ es de Fotografia 3 (continued on next page) P. Alves et al. Expert Systems With Applications 255 (2024) 124812 34 Table B2 (continued) User # Subgroup POI ID POI name Rating/Interest 19 C9Sg 683 Passadiços do Paiva 5 19 C9Sg 486 Mata Nacional do Buçaco 3 19 C9Sg 33 Arouca Geopark 4 26 C7Sg 57 Bom Jesus do Monte 5 26 C7Sg 652 Pal´ acio da Bolsa 5 26 C7Sg 885 S´ e Catedral do Porto 5 26 C7Sg 382 Igreja de S˜ ao Francisco 2 26 C7Sg 486 Mata Nacional do Buçaco 3 26 C7Sg 519 MontanhaParque da Penha 5 26 C7Sg 283 Excurs˜ ao de meio dia ao Pal´ acio Bussaco, jardins e floresta 4 26 C7Sg 588 Museu FC Porto 3 26 C7Sg 624 O Navio Gil Eannes 5 26 C7Sg 174 Centro Portuguˆ es de Fotografia 4 24 C5Sg 652 Pal´ acio da Bolsa 4 24 C5Sg 885 S´ e Catedral do Porto 4 24 C5Sg 382 Igreja de S˜ ao Francisco 4 24 C5Sg 170 Centro Hist´ orico de Guimar˜ aes 4 24 C5Sg 588 Museu FC Porto 5 24 C5Sg 624 O Navio Gil Eannes 2 24 C5Sg 174 Centro Portuguˆ es de Fotografia 2 24 C5Sg 690 Pateira de Fermentelos 4 24 C5Sg 708 Poço Negro 4 24 C5Sg 692 Pedra Bolideira 1 The POI names are in Portuguese. Appendix C Table C1 Tourist attraction types considered for each tourism category. Attraction type Tourism category A12 Practice aquatic sports (e.g., sailing, canoeing, diving, jet skiing) Adrenaline activities A24 Ride a bike Adrenaline activities A29 Do motorsports (e.g., karting, motocross) Adrenaline activities A37 Observe sub-aquatic environments/marine life (e.g., snorkeling, submarine) Adrenaline activities A46 Practice climbing or bungee jumping Adrenaline activities A63 Do air sports (e.g., parachute jump, skydiving, gliding) Adrenaline activities A68 Ski Adrenaline activities A3 Watch a religious celebration Cultural heritage A4 Visit the historic cities/villages of the destination Cultural heritage A7 Visit archaeological sites/ruins Cultural heritage A20 Visit monuments (e.g. churches, cathedrals, castles, fortresses, monasteries, palaces, etc.) Cultural heritage A23 To enjoy/buy local handicrafts Cultural heritage A38 Visit large man-made constructions (e.g., bridges, tunnels, mines) Cultural heritage A18 Taste typical local dishes Gastronomy events A40 Participate in a gastronomy tour (typical and/or gourmet dishes, wine tasting) Gastronomy events A14 Undergo health and wellness treatments (e.g., hydrotherapy centers, mineral water resorts) Health &well-being A26 Attend gyms/fitness centers Health &well-being A28 Go to a SPA/beauty center Health &well-being A32 Visit museums of historical themes Museums, boat trips &viewpoints A33 Visit museums of scientific themes (e.g., planetarium, paleontology) Museums, boat trips &viewpoints A34 Visit viewpoints of natural landscape Museums, boat trips &viewpoints A35 Visit viewpoints of urban landscape Museums, boat trips &viewpoints A49 Take boat trips to know the destination’s coast Museums, boat trips &viewpoints A50 Take boat trips for the historical value of the route Museums, boat trips &viewpoints A51 Take boat trips for the pleasure of boating Museums, boat trips &viewpoints A2 Watch a natural phenomenon (e.g., volcanic eruption or northern lights) Natural phenomena A6 Visit caves/caverns/volcanoes Natural phenomena A8 Attend cultural activities/artistic performances Party, Music &Nightlife A9 Go to the disco/nightclub Party, Music &Nightlife A16 Attend a typical celebration of the destination (e.g., popular celebrations, carnivals, fireworks) Party, Music &Nightlife A17 Go to a film festival Party, Music &Nightlife A39 Go to a thematic parade (e.g., military, electronic music) Party, Music &Nightlife A43 Go to a music festival/concert Party, Music &Nightlife A44 Go to a dance/ballet festival Party, Music &Nightlife A45 Go to balls (dancing) Party, Music &Nightlife A48 Go to a live music bar/place Party, Music &Nightlife A67 Assist an opera/theater Party, Music &Nightlife A53 Play ball sports (e.g., football, handball, volleyball, tennis) Sports &Games A55 Play at the casino Sports &Games A56 Assist to a sporting competition (e.g., watch a football game from a club of that country) Sports &Games A58 Hunt/fish Sports &Games (continued on next page) P. Alves et al. Expert Systems With Applications 255 (2024) 124812 35 Table C1 (continued) Attraction type Tourism category A59 Participate in an escape game Sports &Games A60 Watch a bullfight Sports &Games A61 Go to the circus Sports &Games A22 Go to the beach (sunbathing/swimming) Sun, water &sand A62 Go on a cruise Sun, water &sand A64 Go to the swimming pool to swim/dive Sun, water &sand A65 Go to the swimming pool to relax Sun, water &sand A66 Have vacation on an island Sun, water &sand A5 Visit an oceanarium Theme &animal parks A13 Go to a theme park (e.g., Disneyland Paris) Theme &animal parks A15 Go to a Zoo Theme &animal parks A19 Visit a botanical garden Theme &animal parks A25 Go to a funfair (e.g., amusements such as Ferris wheel, bumper cars, etc.) Theme &animal parks A27 Go to a water park Theme &animal parks A10 Appreciate natural landscapes Wild nature activities A11 Do hiking/mountaineering Wild nature activities A21 Visit a beach for its natural beauty Wild nature activities A36 Visit nature or wildlife reserves Wild nature activities A41 Walk in the forest/woods Wild nature activities A42 Take a walk along the river/sea coast Wild nature activities A47 Visit mountain areas/gorges Wild nature activities A54 Do a safari Wild nature activities A57 Ride a horse Wild nature activities Table C2 Default d-means algorithm and application configuration parameters. Parameter Default value Minimum tourists’similarity 0.80 Apriori rules Confidence 95% Apriori rules Support 95% Rules check frequency every 24h Minimum tourists to divide main excursion 6 Minimum tourists in a subgroup 3 Minimum tourists in a cluster to search rules 50 Minimum tourists in a subgroup to match a rule for POI inclusion/exclusion 50% Minimum POI to recommend in a list 10 Minimum rating to exclude a POI ≤2 Threshold to consider a category as the Top-3 0.49 Minimum travel preference value to consider that preference ≥0.50 Minimum travel concern value to consider that concern ≥0.50 Weather API check frequency every 24h Appendix D. Supplementary data Supplementary data to this article can be found online at https://doi.org/10.1016/j.eswa.2024.124812. References Abernathy, A., & Celebi, M. E. (2022). The incremental online k-means clustering algorithm and its application to color quantization. Expert Systems with Applications, 207, Article 117927. Abolghasemi, R., Engelstad, P., Herrera-Viedma, E., & Yazidi, A. (2022). A personalityaware group recommendation system based on pairwise preferences. Information Sciences, 595, 1–17. Ackermann, M. R., M¨ artens, M., Raupach, C., Swierkot, K., Lammersen, C., & Sohler, C. (2012). Streamkm++a clustering algorithm for data streams. Journal of Experimental Algorithmics (JEA), 17, 2.1-2.30. Aggarwal, C. C., Philip, S. Y., Han, J., & Wang, J. (2003). A framework for clustering evolving data streams. Paper presented at the Proceedings 2003 VLDB conference. Altman, N., & Krzywinski, M. (2018). The curse (s) of dimensionality. Nature Methods, 15 (6), 399–400. ´ Alvarez M´ arquez, J. O., & Ziegler, J. (2016). Hootle+: A group recommender system supporting preference negotiation. Paper presented at the Collaboration and Technology: 22nd International Conference, CRIWG 2016, Kanazawa, Japan, September 14-16, 2016, Proceedings 22. Alves, P., Carneiro, J., Marreiros, G., & Novais, P. (2019). Modeling a mobile group recommender system for tourism with intelligent agents and gamification. Paper presented at the International Conference on Hybrid Artificial Intelligence Systems. Alves, P., Gomes, D., Rodrigues, C., Carneiro, J., Novais, P., & Marreiros, G. (2022). Grouplanner: A Group Recommender System for Tourism with Multi-agent MicroServices. Paper presented at the Advances in Practical Applications of Agents, Multi-Agent Systems. Paper presented at the Advances in Practical Applications of Agents, Multi-Agent Systems, and Complex Systems Simulation. The PAAMS Collection: 20th International Conference, PAAMS 2022, L’Aquila, Italy, July 13–15, 2022, Proceedings. Alves, P., Martins, A., Novais, P., & Marreiros, G. (2023). Improving Group Recommendations using Personality, Dynamic Clustering and Multi-Agent MicroServices. Paper presented at the Proceedings of the 17th ACM Conference on Recommender Systems. Alves, P., Martins, H., Saraiva, P., Carneiro, J., Novais, P., & Marreiros, G. (2023). Group Recommender Systems for Tourism: How does Personality predict Preferences for Attractions, Travel Motivations, Preferences and Concerns? User Modeling and UserAdapted Interaction.https://doi.org/10.1007/s11257-023-09361-2 Alves, P., Saraiva, P., Carneiro, J., Campos, P., Martins, H., Novais, P., &Marreiros, G. (2020). Modeling Tourists’Personality in Recommender Systems: How Does Personality Influence Preferences for Tourist Attractions? Paper presented at the Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization. Alves, P., Trindade, J., Monteiro, G., Saraiva, P., Campos, P., Marreiros, G., &Novais, P. (2024). “You Want to Play a Game?”Detecting Personality Traits with Mobile Minigames. TechRxiv. Retrieved from 10.36227/techrxiv.171177596.69544930/v1. Amato, A. (2023). Recommender Systems in the Museum Sector: An Overview. Paper presented at the Advanced Information Networking and Applications: Proceedings of P. Alves et al. Expert Systems With Applications 255 (2024) 124812 36 the 37th International Conference on Advanced Information Networking and Applications (AINA-2023), Volume 3. Ardissono, L., Goy, A., Petrone, G., Segnan, M., & Torasso, P. (2003). Intrigue: Personalized recommendation of tourist attractions for desktop and hand held devices. Applied Artificial Intelligence, 17(8–9), 687–714. Arthur, D., & Vassilvitskii, S. (2007). k-means++: The advantages of careful seeding. Paper presented at the Soda. Ayala-G´ omez, F., Dar´ oczy, B. Z., Mathioudakis, M., Benczúr, A., &Gionis, A. (2017). Where could we go? Recommendations for groups in location-based social networks. Barkan, O., Bogina, V., Gurevitch, L., Asher, Y., & Koenigstein, N. (2024). A Counterfactual Framework for Learning and Evaluating Explanations for Recommender Systems. Paper presented at the Proceedings of the ACM on Web Conference 2024. Batet, M., Moreno, A., S´ anchez, D., Isern, D., & Valls, A. (2012). Turist@: Agent-based personalised recommendation of tourist activities. Expert Systems with Applications, 39(8), 7319–7329. Bekkerman, P., Kraus, S., &Ricci, F. (2006). Applying cooperative negotiation methodology to group recommendation problem. Paper presented at the Proceedings of Workshop on Recommender Systems in 17th European Conference on Artificial Intelligence (ECAI 2006). Bengali, V., &Veldt, N. (2023). Faster Approximation Algorithms for Parameterized Graph Clustering and Edge Labeling. Paper presented at the ACM International Conference on Information and Knowledge Management (CIKM ’23), Birmingham, United Kingdom. Bi, X., Qu, A., Wang, J., & Shen, X. (2017). A group-specific recommender system. Journal of the American Statistical Association, 112(519), 1344–1353. Blanco, J. M., Ge, M., &Pitner, T. (2023). An Adaptive Filter for Preference Fine-Tuning in Recommender Systems. Paper presented at the Web Information Systems and Technologies: 16th International Conference, WEBIST 2020, November 3–5, 2020, and 17th International Conference, WEBIST 2021, October 26–28, 2021, Virtual Events, Revised Selected Papers. Boratto, L., &Carta, S. (2010). State-of-the-art in group recommendation and new approaches for automatic identification of groups. In Information retrieval and mining in distributed environments (pp. 1-20): Springer. Boratto, L., & Carta, S. (2015). ART: Group recommendation approaches for automatically detected groups. International Journal of Machine Learning and Cybernetics, 6(6), 953–980. Borr` as, J., Moreno, A., & Valls, A. (2014). Intelligent tourism recommender systems: A survey. Expert systems with applications, 41(16), 7370–7389. Cantador, I., & Fern´ andez-Tobías, I. (2014). On the exploitation of user personality in recommender systems. Paper presented at the CEUR Workshop Proceedings. Carneiro, J., Andrade, R., Alves, P., Conceiç˜ ao, L., Novais, P., & Marreiros, G. (2020). A consensus-based group decision support system using a multi-agent MicroServices approach. Paper presented at the Proceedings of the 19th International Conference on Autonomous Agents and MultiAgent Systems. Carneiro, J., Martinho, D., Marreiros, G., & Novais, P. (2018). Arguing with Behavior Influence: A Model for Web-based Group Decision Support Systems. International Journal of Information Technology &Decision Making, 1–37. Chandramouli, B., Levandoski, J. J., Eldawy, A., & Mokbel, M. F. (2011). Streamrec: A real-time recommender system. Paper presented at the Proceedings of the 2011 ACM SIGMOD International Conference on Management of data. Chen, Y., &Pu, P. (2012). CoFeel: Using emotions for social interaction in group recommender systems. Paper presented at the First International Workshop on Recommendation Technologies for Lifestyle Change (LIFESTYLE 2012). Cohen, J. (2013). Statistical power analysis for the behavioral sciences: Academic press. Costa, P. T., &MacCrae, R. R. (1992). Revised NEO personality inventory (NEO PI-R) and NEO five-factor inventory (NEO-FFI): Professional manual: Psychological Assessment Resources, Incorporated. Dara, S., Chowdary, C. R., & Kumar, C. (2020). A survey on group recommender systems. Journal of intelligent Information Systems, 54(2), 271–295. De Campos, L. M., Fern´ andez-Luna, J. M., & Huete, J. F. (2024). An explainable contentbased approach for recommender systems: A case study in journal recommendation for paper submission. User Modeling and User-Adapted Interaction, 1–35. Delic, A., Emamgholizadeh, H., Nguyen, T. N., &Ricci, F. (2024). CHARM: a Group Decision Making Support Chatbot. Paper presented at the Companion Proceedings of the 29th International Conference on Intelligent User Interfaces. Delic, A., &Masthoff, J. (2018). Group Recommender Systems. Paper presented at the Proceedings of the 26th Conference on User Modeling, Adaptation and Personalization. Delic, A., Neidhardt, J., Rook, L., Werthner, H., &Zanker, M. (2017). Researching individual satisfaction with group decisions in tourism: experimental evidence. Paper presented at the Information and Communication Technologies in Tourism 2017: Proceedings of the International Conference in Rome, Italy, January 24-26, 2017. Deli´ c, A., Nguyen, T. N., & Tkalˇ ciˇ c, M. (2020). Group decision-making and designing group recommender systems. Handbook of e-Tourism, 1–23. Digman, J. M. (1990). Personality structure: Emergence of the five-factor model. Annual Review of Psychology, 41(1), 417–440. Djebali, S., Gabot, Q., &Guerard, G. (2023). Hierarchical Clustering and Measure for Tourism Profiling. Paper presented at the Web and Big Data: 6th International Joint Conference, APWeb-WAIM 2022, Nanjing, China, November 25–27, 2022, Proceedings, Part II. El Majjodi, A., Elahi, M., El Ioini, N., &Trattner, C. (2020). Towards Generating Personalized Country Recommendation. Paper presented at the Adjunct Publication of the 28th ACM Conference on User Modeling, Adaptation and Personalization. Emamgholizadeh, H., Deli´ c, A., &Ricci, F. (2024). Preference Management in a Single User Group Recommender System. Paper presented at the ENTER e-Tourism Conference. Ezugwu, A. E., Ikotun, A. M., Oyelade, O. O., Abualigah, L., Agushaka, J. O., Eke, C. I., & Akinyelu, A. A. (2022). A comprehensive survey of clustering algorithms: State-ofthe-art machine learning applications, taxonomy, challenges, and future research prospects. Engineering Applications of Artificial Intelligence, 110, Article 104743. Feldman, R., & Sanger, J. (2007). The text mining handbook: Advanced approaches in analyzing unstructured data. Cambridge University Press. Ferwerda, B., Schedl, M., &Tkalcic, M. (2015). Predicting personality traits with instagram pictures. Paper presented at the Proceedings of the 3rd Workshop on Emotions and Personality in Personalized Systems 2015. Field, A. (2013). Discovering statistics using IBM SPSS statistics: sage. Finnerty, A. N., Lepri, B., & Pianesi, F. (2016). Acquisition of personality. In Emotions and Personality in Personalized Services (pp. 81–99). Springer. Forsyth, D. R. (2018). Group dynamics: Cengage Learning. Garcia, I., Sebastia, L., & Onaindia, E. (2011). On the design of individual and group recommender systems for tourism. Expert Systems with Applications, 38(6), 7683–7692. Garcia, I., Sebastia, L., Onaindia, E., & Guzman, C. (2009). A group recommender system for tourist activities. Paper Presented at the International Conference on electronic commerce and web. technologies. Gavalas, D., Konstantopoulos, C., Mastakas, K., & Pantziou, G. (2014). Mobile recommender systems in tourism. Journal of Network and Computer Applications, 39, 319–333. Goldberg, L. R. (1990). An alternative“description of personality”: The big-five factor structure. Journal of Personality and Social Psychology, 59(6), 1216. Gravetter, F. J., Wallnau, L. B., Forzano, L.-A.-B., & Witnauer, J. E. (2020). Essentials of statistics for the behavioral sciences. Cengage Learning. Gretzel, U., Mitsche, N., Hwang, Y.-H., & Fesenmaier, D. R. (2006). Travel personality testing for destination recommendation systems (pp. 121–136). Oxfordshire: Destin Recomm Syst Behav Found Appl. CABI. Guzzi, F., Ricci, F., &Burke, R. (2011). Interactive multi-party critiquing for group recommendation. Paper presented at the Proceedings of the fifth ACM conference on Recommender systems. He, B., He, X., Zhang, R., Zhang, Y., Tang, R., &Ma, C. (2023). Dynamic Embedding Size Search with Minimum Regret for Streaming Recommender System. Paper presented at the Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. Ismailoglu, F. (2022). Aggregating user preferences in group recommender systems: A crowdsourcing approach. Decision Support Systems, 152, Article 113663. Jameson, A. (2004). More than the sum of its members: challenges for group recommender systems. Paper presented at the Proceedings of the working conference on Advanced visual interfaces. Jameson, A., Baldes, S., &Kleinbauer, T. (2003). Enhancing mutual awareness in group recommender systems. Paper presented at the Proceedings of the IJCAI. Kaufman, L., & Rousseeuw, P. J. (2009). Finding groups in data: An introduction to cluster analysis. John Wiley &Sons. Khazaei, E., & Alimohammadi, A. (2019). Context-aware group-oriented location recommendation in location-based social networks. ISPRS International Journal of Geo-Information, 8(9), 406. Kuo, R., & Li, S.-S. (2023). Applying particle swarm optimization algorithm-based collaborative filtering recommender system considering rating and review. Applied Soft Computing, 110038. Lam, D., & Wunsch, D. C. (2014). Clustering. Academic Press Library in Signal Processing, 1, 1115–1149. Liu, H., Zhao, X., Wang, C., Liu, X., & Tang, J. (2020). Automated embedding size search in deep recommender systems. Paper presented at the Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. Lorenzi, F., Loh, S., & Abel, M. (2011). PersonalTour: A recommender system for travel packages. Paper presented at the 2011 IEEE/WIC/ACM International Conferences on Web Intelligence and Intelligent Agent Technology. Lorenzi, F., Loh, S., & Abel, M. (2012). Personal Tour: A Multi-Agent Recommender System of Travel Packages. In Global Hospitality and Tourism Management Technologies (pp. 178–189). IGI Global. Lourenço, A., Meira, J., & Marreiros, G. (2023). Online adaptive learning for out-ofround railway wheels detection. Paper presented at the Proceedings of the 38th ACM/ SIGAPP Symposium on Applied Computing. Mai, S. T., Assent, I., & Storgaard, M. (2016). AnyDBC: An efficient anytime densitybased clustering algorithm for very large complex datasets. Paper presented at the Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining. Marˆ oco, J. (2010). An´ alise de Equaç˜ oes Estruturais: Fundamentos te´ oricos, software & Aplicaç˜ oes: ReportNumber, Lda. Marques, G., Respício, A., & Afonso, A. P. (2016). A mobile recommendation system supporting group collaborative decision making. Procedia Computer Science, 96, 560–567. Martijn, M., Conati, C., & Verbert, K. (2022). “Knowing me, knowing you”: Personalized explanations for a music recommender system. User Modeling and User-Adapted Interaction, 1–38. Massimo, D., & Ricci, F. (2021). Next-POI recommendations matching user’s visit behaviour. Paper presented at the Information and Communication Technologies in Tourism 2021: Proceedings of the ENTER 2021 eTourism Conference. Masthoff, J. (2015). Group recommender systems: Aggregation, satisfaction and group attributes. In In recommender systems handbook (pp. 743–776). Springer. McCarthy, K., McGinty, L., Smyth, B., & Salam´ o, M. (2006). Social interaction in the cats group recommender. Paper presented at the Workshop on the Social Navigation and Community Based Adaptation Technologies. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 37 McCarthy, K., Salam´ o, M., Coyle, L., McGinty, L., Smyth, B., & Nixon, P. (2006). Cats: A synchronous approach to collaborative group recommendation. Paper presented at the Florida Artificial Intelligence Research Society Conference (FLAIRS). McCrae, R. R., & Costa, P. T., Jr (1997). Personality trait structure as a human universal. American Psychologist, 52(5), 509. Michaelsen, L. K., Watson, W. E., & Black, R. H. (1989). A realistic test of individual versus group consensus decision making. Journal of Applied Psychology, 74(5), 834. Nadeem, M., Wasid, M., Nadeem, M., Alam, M. T., Sohail, S. S., Ubaid, S., & Shahab, S. (2022). Performance Comparison of Randomized and Non-Randomized Learning Algorithms based Recommender Systems. International Journal of Next-Generation Computing, 13(3). Najafian, S., Herzog, D., Qiu, S., Inel, O., & Tintarev, N. (2020). You do not decide for me! evaluating explainable group aggregation strategies for tourism. Paper presented at the Proceedings of the 31st ACM Conference on Hypertext and Social Media. Neidhardt, J., Schuster, R., Seyfang, L., & Werthner, H. (2014). Eliciting the users’ unknown preferences. Paper presented at the Proceedings of the 8th ACM Conference on Recommender systems. Neidhardt, J., Seyfang, L., Schuster, R., & Werthner, H. (2015). A picture-based approach to recommender systems. Information Technology &Tourism, 15(1), 49–69. Neto, J., Morais, A. J., Gonçalves, R., & Coelho, A. L. (2022a). Context-Based Multi-Agent Recommender System, Supported on IoT, for Guiding the Occupants of a Building in Case of a Fire. Electronics, 11(21), 3466. Neto, J., Morais, A. J., Gonçalves, R., & Coelho, A. L. (2022b). Multi-Agent-Based Recommender Systems: A Literature Review. Information and Communication Technology. Nguyen, T. N., & Ricci, F. (2018). A chat-based group recommender system for tourism. Information Technology &Tourism, 18(1–4), 5–28. Nozari, R. B., & Koohi, H. (2020). A novel group recommender system based on members’influence and leader impact. Knowledge-Based Systems, 205, Article 106296. Ntoutsi, E., Stefanidis, K., Nørvåg, K., & Kriegel, H.-P. (2012). Fast group recommendations by applying user clustering. Paper presented at the Conceptual Modeling: 31st International Conference ER 2012, Florence, Italy, October 15-18, 2012. Proceedings 31. Nunes, M. A. S. N., Cerri, S. A., & Blanc, N. (2008). Improving recommendations by using personality traits in user profiles. Paper presented at the International Conferences on Knowledge Management and New Media Technology. Park, M.-H., Park, H.-S., & Cho, S.-B. (2008). Restaurant recommendation for group of people in mobile environments using probabilistic multi-criteria decision making. Paper presented at the Computer-Human Interaction: 8th Asia-Pacific Conference, APCHI 2008 Seoul, Korea, July 6-9, 2008 Proceedings 8. Pedregon, C. A., Farley, R. L., Davis, A., Wood, J. M., & Clark, R. D. (2012). Social desirability, personality questionnaires, and the “better than average”effect. Personality and Individual Differences, 52(2), 213–217. Pei, F., Gao, Y., Yan, A., Zhou, M., &Wu, J. (2024). Conflict elimination based on opinion dynamics in fuzzy group decision-making. Expert systems with applications, 124308. Pham, D. T., Dimov, S. S., & Nguyen, C. (2004). An incremental K-means algorithm. Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science, 218(7), 783–795. Ravi, L., Devarajan, M., Sangaiah, A. K., Wang, L., & Subramaniyaswamy, V. (2021). An intelligent location recommender system utilising multi-agent induced cognitive behavioural model. Enterprise Information Systems, 15(10), 1376–1394. Rawlings, D., & Ciancarelli, V. (1997). Music preference and the five-factor model of the NEO Personality Inventory. Psychology of Music, 25(2), 120–132. Ricci, F. (2022). Recommender systems in tourism. In Handbook of e-Tourism (pp. 457–474). Springer. Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. Journal of Computational and Applied Mathematics, 20, 53–65. Saxena, A., Prasad, M., Gupta, A., Bharill, N., Patel, O. P., Tiwari, A., & Lin, C.-T. (2017). A review of clustering techniques and developments. Neurocomputing, 267, 664–681. Sertkan, M., Neidhardt, J., & Werthner, H. (2019). What is the “Personality”of a tourism destination? Information Technology &Tourism, 21(1), 105–133. Sertkan, M., Neidhardt, J., &Werthner, H. (2020). PicTouRe-A Picture-Based Tourism Recommender. Paper presented at the Fourteenth ACM Conference on Recommender Systems. Shams, S., Anderson, D., &Leith, D. (2021). Cluster-based bandits: Fast cold-start for recommender system new users. Paper presented at the Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. Sharma, D., Banwala, E., Elouaghzani, I., &Katarya, R. (2023). Analysis of Personalized Tourism Recommender Systems (2516-2314). Retrieved from. Shaw, M. E. (1932). A comparison of individuals and small groups in the rational solution of complex problems. The American Journal of Psychology, 44(3), 491–504. Silva, D. L. C. d. (2023). Development of an Ontology of Tourist Attractions for Recommending Points of Interest in a Group Recommender System for Tourism. (Master), Polit´ ecnico do Porto. Silva, J. A., Faria, E. R., Barros, R. C., Hruschka, E. R., Carvalho, A. C. D., & Gama, J. (2013). Data stream clustering: A survey. ACM Computing Surveys (CSUR), 46(1), 1–31. Sojahrood, Z. B., Taleai, M., & Cheng, H. (2023). Hybrid POI Group Recommender System based on group type in LBSN. Expert Systems with Applications, 119681. Sposito, V., Hand, M., & Skarpness, B. (1983). On the efficiency of using the sample kurtosis in selecting optimal lpestimators. Communications in Statistics-Simulation and Computation, 12(3), 265–272. Su, L., Cheng, J., & Huang, Y. (2021). How do group size and group familiarity influence tourist satisfaction? The mediating Role of perceived value. Journal of Travel Research, 60(8), 1821–1840. Thomas, K. W. (2008). Thomas-kilmann conflict mode. TKI Profile and Interpretive Report, 1(11). Tkalcic, M., & Chen, L. (2015). Personality and recommender systems. In Recommender systems handbook (pp. 715–739). Springer. Tondello, G. F., Orji, R., &Nacke, L. E. (2017). Recommender systems for personalized gamification. Paper presented at the Adjunct Publication of the 25th Conference on User Modeling, Adaptation and Personalization. Tunguz, B. (2018). Big Five Personality Test. Retrieved from https://www.kaggle.com/ datasets/tunguz/big-five-personality-test/data. Vathis, N., Konstantopoulos, C., Pantziou, G., & Gavalas, D. (2023). The Vacation Planning Problem: A multi-level clustering-based metaheuristic approach. Computers &Operations Research, 150, Article 106083. W. Collier, R., O’Neill, E., Lillis, D., &O’Hare, G. (2019). MAMS: Multi-Agent MicroServices. Paper presented at the Companion Proceedings of The 2019 World Wide Web Conference. Wallinger, S., Grundner, L., Majic, I., &Lampoltshammer, T. J. (2023). Agent-Based Modelling for Sustainable Tourism. Paper presented at the Information and Communication Technologies in Tourism 2023: Proceedings of the ENTER 2023 eTourism Conference, January 18-20, 2023. Wooldridge, M., & Jennings, N. R. (1995). Intelligent agents: Theory and practice. The Knowledge Engineering Review, 10(2), 115–152. Xu, F., Tian, F., Buhalis, D., Weber, J., & Zhang, H. (2016). Tourists as mobile gamers: Gamification for tourism marketing. Journal of Travel &Tourism Marketing, 33(8), 1124–1142. Yalcin, E., Ismailoglu, F., & Bilge, A. (2021). An entropy empowered hybridized aggregation technique for group recommender systems. Expert Systems with Applications, 166, Article 114111. Yee, N., Ducheneaut, N., Nelson, L., &Likarish, P. (2011). Introverted elves &conscientious gnomes: the expression of personality in world of warcraft. Paper presented at the Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. Zhang, H., Sun, Y., Zhao, M., Chow, T. W., & Wu, Q. J. (2019). Bridging user interest to item content for recommender systems: An optimization model. IEEE Transactions on Cybernetics, 50(10), 4268–4280. Patrícia Alves is a PhD. Student at University of Minho (UMinho) and researcher at the ALGORITMI Centre and GECAD, ISEP, IPP, Portugal. She’s also Guest Assistant Professor at ISEP. She has a master’s in Informatics and Computing Engineering from the Faculty of Engineering, University of Porto (UP); and a degree in Chemistry (Pre-Bologna) from the Faculty of Sciences, UP. She was awarded the “IBM Demonstration Award”1 st prize at PAAMS’22 and the 2 nd best paper prize at Pro-VE’12 conferences. She’s supervising several undergraduate and master’s students. Her main interests are Artificial Intelligence, Recommender Systems, Affective Computing, Gamification and AR Pervasive Games. Andr´ e Martins has a master’s in Software Engineering from ISEP and has a degree in Informatics Engineering also from ISEP. He is currently working as a research fellow at the Research Group on Intelligent Engineering and Computing for Advanced Innovation and Development (GECAD, ISEP). His main interests include Software Engineering (backend), Web Development and Artificial Intelligence. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 38 Francisco Negr˜ ao has a master’s in Software Engineering and a degree in Computer Engineering from ISEP. He is currently working as a research fellow at GECAD, ISEP. Throughout his studies, Francisco has shown interest in Software Engineering, standing out for his passion for innovation and development of advanced technological solutions. He is dedicated to improving his technical skills and knowledge in this area, keeping up to date with the latest trends and practices. Paulo Novais is a Full Professor of Computer Science at the Department of Informatics, School of Engineering, UMinho and researcher at the ALGORITMI Centre where he’s the leader of ISLab research group (Synthetic Intelligence lab). He’s the coordinator of the Portuguese Intelligent Systems Associate Laboratory (LASI). President of General Assembly and former president of the Portuguese Association for Artificial Intelligence, Senior member of the IEEE, Member of the IFIP-TC12 Artificial Intelligence and of the executive committee of IBERAMIA. His main research objective is to make systems a little more smart, reliable and sensitive to human presence and interaction. Ana Almeida is an Associated Professor at ISEP, IPP, President of Technical Scientific Council, and researcher at GECAD, ISEP. She has a PhD. in Systems &Industrial Engineering, M.Sc. in Electrical and Computers Engineering, specialization in Informatics Engineering and Informatics Engineering Degree from ISEP. She is member of the External Evaluation Board of the Portuguese Agency for Evaluation and Accreditation of Higher Education Study Cycles, in the fields of Computer Engineering and Sciences. Her main research interests include AI applications, Intelligent Systems, Decision Support, BI and Affective Computing. Goreti Marreiros is a Coordinator Teacher with Habilitation at ISEP, GECAD’s Director and Senior researcher, Subdirector of the master’s in Engineering and Artificial Intelligence and Member of ISEP’s Scientific Council. She’s vice-president of the Portuguese Association of Artificial Intelligence. She has Degree and PhD. in Informatics –Artificial Intelligence by UMinho. Responsible for courses on AI at the graduate and postgraduate levels. Participates/ed as Principal Investigator and Researcher in European and National projects. Works in the areas of Affective Computing, Ambient Intelligence and Intelligent Decision Support Systems. Her research work interest are Group Decision Support, Industry 4.0 and Health Care and wellbeing. P. Alves et al. Expert Systems With Applications 255 (2024) 124812 39