An ontology-based data integration approach for web analytics in e-commerce

An ontology-based data integration approach for web analytics in e-commerce María del Mar Roldán García, José García-Nieto ∗, José F. Aldana-Montes 1 Dept. de Lenguajes y Ciencias de la Computación, University of Málaga, ETSI Informática, Campus de Teatinos, Málaga - 29071, Spain Keywords: Semantic model Ontology E-commerce Web analytics a b s t r a c t Web analytics has emerged as one of the most important activities in e-commerce, since it allows companies and e-merchants to track the behavior of customers when visiting their web sites. There exist a series of tools for web analytics that are used not only for tracking and measuring web traffic, but also for analyzing the commercial activity. However, most of these tools focus on low level web attributes and metrics, making other sophisticated functionalities and analyses only available for commercial (non-free) versions. In this context, the SME-Ecompass European initiative aims at providing e-commerce SMEs with accessible tools for high level web analytics. These software facilities should use different sources of data coming from digital footprints allocated in e-shops, to fuse them together in a coherent way, and to make them available for advanced data mining procedures. This motivated us to propose in this work an ontology-based approach to collect, integrate and store web analytics data, from many sources of popular and commercial digital footprints. As article’s main impact, we obtain enriched and semantically annotated data that is used to properly train an intelligent system, involving data mining procedures, for the analysis of customer behavior in real e-commerce sites. In concrete, for the validation of our semantic approach, we have captured and integrated data from Google Analytics and Piwik digital footprints allocated in 15 e-shops of different commercial sectors and countries (UK, Spain, Greece and Germany), throughout several months of activity. The obtained results show different perspectives in customer’s behavior analysis that go one step beyond the most popular web analytics tools in the current market. 1. Introduction In the last few years, web analytics has emerged as one of the most important activities in e-commerce, since it allows companies and e-merchants to track the behavior of customers when visiting their e-shop sites. Web analytic applications can also help companies to measure the results of traditional print or broadcast advertising campaigns. Web analytics procedure is based on measuring a visitor’s behavior once on a given e-shop site, which includes its drivers and conversions (to actual customer). These data are typically compared against key performance indicators and used to improve a website or marketing campaign’s audience response. ∗Corresponding author. E-mail addresses: [email protected] (M.d.M.R. García), [email protected] , In the current market, there exist a series of tools for web analytics, such as: Google Analytics, Piwik, Clicky, and StatCounter; that are widely used not only for tracking and measuring web traffic, but also for analyzing the commercial activity, hence to improve the effectiveness of a website. However, these tools often focus on low level and limited sets of web metrics and attributes, without the possibility of providing specialized analyses. In most of cases, high level web metrics and sophisticated functionalities are available only for commercial (non-free) versions, which are rarely accessible by SMEs or individual e-merchants. In this context, the SME-Ecompass European initiative 2 aims at providing e-commerce SMEs with accessible tools for high level web analytics. These software facilities use the different sources of data coming from different digital footprints allocated in e-shops. However, integrating data from multiple heterogeneous sources entails dealing with different data models, schema and query languages. Therefore, there is a clear demand of integrative proce2 SME-Ecompass FP7 European initiative http://www.sme-ecompass.eu/ jmgarcianie[email protected] (J. García-Nieto), jf[email protected] (J.F. Aldana-Montes). 1 This work is partially funded by FP7 EU project SME E-COMPASS under Grant No: 315637. It is also partially funded by Grants TIN2014-58304 (Spanish Ministry of Sciences and Innovation) and Regional projects P11-TIC-7529/P12-TIC-1519. Authors thanks to involved e-shops to kindly offer web tracking data for testing and validation. dures for providing the advanced data mining algorithms with a uniform access to multiple heterogeneous web data sources. The main hypothesis in this work is: (H1) an ontology-based integration approach will help us to collect, fuse the data together in a coherent way, and store web analytics data, from many sources of popular and commercial digital footprints . As a result, (H2) we will obtain enriched and semantically annotated data that will be able to train data mining procedures for advanced analysis of customer behavior in real e-commerce sites . This motivated us to propose a semantic approach that uses an ontology as a mediated schema for the representation and consolidation in a knowledge base of the tracking data from web source’s semantics. Semantic web ontologies become a key technology for intelligent knowledge processing, providing a framework for sharing conceptual models about a domain. Semantic mappings between the source schema and the ontology are then defined and used to transform the original data to RDF (Resource Description Framework) 3 . This way, data from heterogeneous sources are stored and integrated inside a single RDF repository, which can be now easily queried by high level algorithms. The goal is to properly feed artificial intelligence procedures capable of deciding how to perform marketing activities, such as: displaying a given advertisement targeted to certain category of clients, or decreasing the price of a product in a given region; then giving rise to sophisticated expert systems for e-commerce applications. The main contributions of this study are summarized as follows: –We have developed a semantic approach for the data integration and consolidation of multiple web analytics data sources. These data are daily accumulated from many heterogeneous digital footprints allocated on actual e-shops. –We have designed and implemented for the first time an OWL (Web Ontology Language) ontology ( Dean & Schreiber, 2004 ) for web analytics. This ontology considers a large and complemented set of attributes and metrics, which have been token from several representative web analytics tools in the market. –To test hypothesis H1, we have captured and integrated data from Google Analytics and Piwik digital footprints allocated in 15 e-shops of different commercial sectors (retail, tourism, electronics, pharmacy, etc.) and countries (UK, Spain, Greece and Germany), throughout several months of activity. The data are integrated following the same (standard) format and stored in a common RDF repository. –To test hypothesis H2, obtained “semantized” data are used to train advanced data mining algorithms to perform customer’s profile analyses. In particular, these algorithms are tested with success in two cases of study to classify the visitor’s behavior and product preference. The remaining of this article is organized as follows. In Section 2 , background and literature overview are presented. Section 3 presents the current state and practices in web analytics for e-commerce. In Section 4 , the semantic approach is described, giving details of the service architecture and the OWL ontology. After this, the validation procedure is reported in Section 5 . Finally, main conclusions and future work are given in Section 6 . 2. Background and related work This section describes the main background concepts. A review of current related works in the specialized literature is carried out to point out their main differences with regards to our approach. 3 RDF in W3C https://www.w3.org/RDF/ 2.1. Background concepts - Ontology. Ontologies provide a formal representation of the real world, shared by a sufficient amount of users, by defining concepts and relationships between them ( Gruber, 1993 ). In computer and information sciences, an ontology defines a set of representational primitives with which to model a domain of knowledge or discourse. These primitives are typically concepts (classes), attributes (properties), class members (class instances) and relationships (property instances). The definitions of the representational primitives include information about their meaning and constraints on their logically consistent application. Ontologies are part of the W3C standards stack for the semantic web, in which they are used to specify standard conceptual vocabularies in which to exchange data between systems, provide services for answering queries, publish reusable knowledge bases, and offer services to facilitate interoperability across multiple, heterogeneous systems and databases. -RDF. Resource Description Framework is a basic ontology language used for representing information about resources on the web ( Staab & Studer, 2009 ). Resources are described in terms of properties and property values using RDF statements. Statements are represented as triples, consisting of a subject, predicate and object. RDF Schema ( Staab & Studer, 2009 ) (RDFS) “semantically extends” RDF to enable us to talk about classes of resources, and the properties that will be used with them. It does this by giving particular meanings to certain RDF properties and resources. RDFS provides the means to describe application specific RDF vocabularies. RDF and RDFS provide basic capabilities for describing vocabularies that describe resources, metadata and ontologies. -SPARQL. It is an RDF query language for ontology models and databases, capable of extracting and manipulating information stored in RDF format. Essentially, SPARQL is a graph-matching query language that can be used to extract knowledge from the model such as the one proposed in this article. Given a data source D, a query consists of a pattern, which is matched against D. The combinations of values resulting from this matching constitute the result of the query ( Pérez, Arenas, & Gutierrez, 2009 ). SPARQL has strong support for querying semi-structured and tagged data, e.g. data with an unpredictable and unreliable structure. SPARQL supports queries to networked, web data sources identified by URIs. In fact, it is a W3C recommendation for RDF data. -OWL. In 2004, the W3C ontology working group ( Dean & Schreiber, 2004 ) proposed OWL as a semantic markup language for publishing and sharing ontologies on the World Wide Web. From a formal point of view, OWL is equivalent to a very expressive description logic where an ontology corresponds to a Tbox ( Gruber, 1993 ). This equivalence allows the language to exploit description logic researcher results. OWL extends RDF and RDFS. When compared to RDF models, OWL adds more vocabulary for describing properties and classes: relations between classes (e.g. disjointedness), cardinality (e.g. “exactly one”), equality, richer typing of properties, characteristics of properties (e.g. symmetry), and enumerated classes ( McGuinness & Harmelen, 2004 ). -OWL-DL. Syntactic variant of the SHOIN (D) description logic ( Haase & Stojanovic, 2005 ) with a different terminology to OWL, which is based on RDFS, hence the support for data values, data types and data type properties. OWL-DL restricts OWL into two distinct ways ( Horrocks & Patel-Schneider, 2003 ): first, some syntactic constructs like recursive descriptions in them are not allowed; second, classes, individuals and properties (respectively concepts, individuals and roles in description logics) must all be disjoint. In this work, we use OWL-DL syntax to formalize the proposed ontology here for our semantic model. A summarized description of basic OWL-DL semantics syntax is shown in Table 1 , where an informal logic syntax is represented (left Table 1 Basic OWL-DL semantic syntax used to formally define the proposed ontology. Descriptions Abstract syntax DL syntax Operators intersection ( C 1 , C 2 , , C n ) C 1 C 2 C n union ( C 1 , C 2 , , C n ) C 1 C 2 C n Restrictions for at least 1 value V from C∃ V.C for all values V from C∀ V.C R is symmetric R ≡R − Class Axioms A partial ( C 1 , C 2 , , C n ) A  C 1 C 2 C n A complete ( C 1 , C 2 , , C n ) A ≡C 1 C 2 C n column) with regards to the corresponding OWL-DL equivalent (right). 2.2. Related works In the last decade, a series of studies have been appearing that semantically model certain domains or sub-domains of knowledge in the context of e-commerce. A first attempt was proposed in Trastour, Bartolini, and Preist (2003) , where a service description language was defined to be used throughout the life-cycle of a business-to-business (B2B) ecommerce interaction. In particular, they focused on DAML+OIL, as it is a sufficiently expressive and flexible service description language to be used not only in advertisements, but also in matchmaking queries, negotiation proposals and agreements. After this, Tamma et al. ( Tamma, Phelps, Dickinson, & Wooldridge, 2005 ) designed an approach to negotiation activities in e-commerce sites. In this work, the negotiation protocol does not need to be hard-coded in agents, but it is represented by an ontology, in terms of an explicit and declarative representation of the negotiation protocol. The ontology is also used to tune agents strategies to the specific protocol used. A special case of e-commerce sites is e-tourism, for which Waralak ( Waralakv, 2008 ) discussed some ontological trends that support the growing domain of online tourism. This study also gave some example concepts of existing e-tourism using ontologies display in graphical model and showed their descriptions in OWL and RDFS syntax. Hepp defined two related ontologies ( Hepp, 2008 ): GoodRelations and Product Ontology. GoodRelations is a standardized vocabulary for e-commerce (product, price, store, and company data) and the Product Ontology is an ontology for describing product types based on Wikipedia. More recently, Gatchalee et al. ( Gatchalee, Li, & Supnithi, 2013 ) proposed an ontology approach to cover the knowledge about the content and architecture of SMEs e-commerce websites. This knowledge is then used as input to a recommendation system for web design, that centered on the structure of e-shops in Thailand as sample group. Finally, Akanbi ( Akanbi, 2014 ) proposed LB2CO, an ontology which combines the framework of IDEF5 & SNAP as an analysis tool, for automated recommendation of product and services. This ontology is used to model a semantic framework for B2C transactions across different business domains that facilitates the interoperability and integration of transactions over the web. As summarized in Table 2 , all these works proposed semantic models focusing on different aspects in the domain of e-commerce, such as: web contents, structure, and life-cycle activities. However, to the best of our knowledge, there is still a lack of works where a semantic model is used to consider web analytic attributes and metrics from multiple and heterogeneous sources of data. This is a critical issue for current e-merchants that we try to cope, for the first time, with our approach. In contrast to other past proposals, we validate our semantic model with real data coming from digital footprints and web scraping methods in a number of actual e-shops. As a result, some of these e-commerce SMEs have put in practice the generated data and analyses, leading them to change and improve their commercial strategies. 3. E-commerce web analytics: current practices Previously to describe our semantic approach, we summarize in this section a series of activities we carried out in the context of SME-Ecompass project, with the aim of shedding light on the actual state of e-commerce companies. In a first phase, we delivered an online survey to e-shop’s owners (of associated chambers to SME-Ecompass project) with different questions with regards to: their company’s profile, commerce activity, current practices when analyzing customer’s and competitor’s behaviors, etc. After that, face-to-face interviews were also conducted with a selection of e-merchants in order to obtain detailed information of their professional experience in e-commerce. A complete report of these questionnaires with the analysis of responses, statistics and conclusions can be found in Garía-Nieto and Roldán (2014) . In concrete, more than 150 e-commerce SMEs completed the online surveys and 20 out of them were interviewed in private sessions. The following conclusions were extracted: – Most of studied companies are micro enterprises with 1 to 5 employees and work in Business to Consumer (B2C) retailing sector. Most of them have a maximum number of 50 0 0 orders per year (2013/2014) and a maximum annual revenue of 10,0 0 0 euros from online sales. Therefore, they are target candidates to be beneficiary of automatic (free or non-expensive) tools to obtain advanced e-commerce analysis. – Concerning visitor’s behavior analyses (see Fig. 1 left), 47% of companies do not use any tool for these kind of tasks. On the contrary, a percentage of 29% declared that they use automatic online tools and 22% make these tasks manually. Of course, most of them ( > 80%) declared to be quite interested on using a service to discover tendencies and common habits in clients. – Interestingly, as shown in Fig. 1 (right), it can be stood out that Google Analytics is the most used tool of interviewed companies (68%), although they also use additional tools like Piwik (16%) and other (16%). Therefore, the set of common metrics emerchants usually analyze are those computed by Google Analytics, e. g., number of visits, average visit time, geo-localization, country, client device, etc. These metrics are usually checked weekly or monthly. In the light of these results, we decided to focus our semantic model on attributes and metrics provided by Google Analytics and Piwik. We selected the former for being the most used analytic tool in the market. However, Google Analytics e-commerce advanced functionality is (to the date) not available for free users. This issue led us to complement our set of attributes with those of Piwik, since this last tool offers free access to advanced ecommerce attributes. Additional metrics for competitor and price monitor are also considered in this work, which are gathered from specific web scrapping processes of SME E-Compass tool. It is worth noting that the proposed ontology is aimed at covering as general as possible web analytic attributes, hence enabling the incorporation of new analytics tools in our semantic approach. 4. Semantic approach One of the main aims of this work is to capture, clean, consolidate and integrate data from different sources of web tracking in Table 2 Related approaches in the state of the art. The target area of application, the used ontology/vocabulary and the postprocessing analysis, and the validation procedure are reported for each work. Approach Target area Ontology/vocabulary Analysis Validation Trastour et al. (2003) Life-cycle in B2B DAML+OIL -No Tamma et al. (2005) Negotiation in e-commerce Negotiation protocol Agents system No Waralakv (2008) E-tourism E-tourism concepts Recommender No Hepp (2008) Products in e-commerce GoodRelations -Academic Gatchalee et al. (2013) E-commerce sites design Website design structure Recommender Academic Akanbi (2014) B2C transactions LB2CO Recommender Academic Proposal Web analytics digital Web analytics Data mining Real footprints (Google/Piwik) Ontology (WAO) Visitor/product world Fig. 1. Current practices in surveyed e-commerce SMEs with regards to the use of automatic tools for web analytics. e-commerce sites. For this reason, we opted to design a semantic approach for sharing and reconciliation, whereby an agreed ontology model is used to archive a common understanding of the domain in which the system operates. In concrete, we have developed an OWL ontology to describe the e-shops main features by following the standard ontology 101 development process ( Natalya, McGuinness, & Deborah, 2001 ) of seven steps: (i) Determine the domain and scope of the ontology . As the starting point, to limit the scope of the ontology, we selected the kind of variables that the data mining algorithms need from Google Analytics and Piwik and also from competitors e-shops, for instance: visitors origin, visitors attributes, purchasing behavior, product and customer details, etc. (ii) Consider reusing existing ontologies . As we examined in Section 2.2 , there are no similar ontologies that have been previously proposed for modeling web tracking data in ecommerce. However, we partially considered two related ontologies: GoodRelations ( Hepp, 2008 ), which is a standardized vocabulary for e-commerce and the Product Ontology ( Hepp, 2008 ), which contextualizes product types based on Wikipedia. (iii) Enumerate important terms in the ontology . Important terms in the ontology were extracted in a previous phase of requirements specification ( Garía-Nieto & Roldán, 2014 ) from the minimum set of variables that are needed. Examples of such terms are: address, visitor, customer, device, browser, Geographical_origin, Number_of_visitors, Conversion_rate, etc. (iv) Define classes and the class hierarchy . From the list of terms, we obtained the ontology classes. Fig. 2 shows the main set of classes in the hierarchy starting from the top class Thing (  ). These main classes are related to other classes and some of them have subclasses. For instance, Analytic_parameters has a series of subclasses, such as: Bounce_rate, Total_revenue, Number_of_returning_visitors , and Number_of_transactions . (v) Define the properties of classes and slots . In order to relate classes and to define attributes, we identified objects and data type properties based on the minimum set of variables previously defined. Examples of object properties are: an e-shop owner is owner of an e-shop, a visitor makes visits, a device has a browser, an IP address belongs to an organization, etc. Examples of data type properties are the title and URL of a page, the first and last name of an e-shop owner, the version of the operating system, the duration of a visit, etc. An object property is defined for each subclass to establish the correct relationship. For example, Page is related to Bounce_rate and Date_of_last_visit; E-shop is related to Number_of_customers . Tables 3–8 , describe in OWL-DL representative subsets of object and data properties of a selection of the main classes. (vi) Define the facets of the slots . This step includes the definition of cardinality constraints and value restrictions. Value restrictions are used in our ontology to specify the data type for the value in each subclass of the Analytic_parameters class. For example, the range of the property hasValue is restricted to float , when the class Bounce_rate is its domain; the range of the property hasValue is restricted to date, when the class Date_of_last_visit is its domain. (vii) Create instances . Instances (individuals in OWL) correspond to the specific data obtained from a specific e-shop. Individuals will be obtained by mapping the data from Google Analytics, Piwik or competitors e-shops to RDF according to the ontology. Individuals can be also used in the ontology to define the exact members of a class. The range of the property hasType is restricted to values: ”ASIN, ”EAN or ”ISBN, when its domain is Article_number. ASIN, EAN and ISBN are then ontology individuals (see Table A.5 for further explanations). 4.1. Ontology model The proposed ontology, called “wao.owl” (Web Analytics Ontology), resulting from the development process described above has a total of 62 classes (groups of individuals sharing the same attributes), 61 object properties (binary relationships between individuals), and 67 data properties (individual attributes), 33 restriction axioms and 3 individuals. The complete ontology is available in WebProtégé repository. 4 4 URL link http://stanford.io/1XHhHzr Fig. 2. General overview of the WAO ontology. Continuous arrows refer to sub class of. Dotted arrows refer to specific properties. Table 3 Analytics_parameters group: object and data properties. Object properties Description logic hasBrowser ∃ hasBrowser.Thing  Analytic_parameters Device   ∀ hasBrowser.Browser hasCity ∃ hasCity.Thing  Analytic_parameters Location Visitor   ∀ hasCity.City hasRegion ∃ hasRegion.Thing  Analytic_parameters Location Visitor   ∀ hasRegion.Region hasCountry ∃ hasCountry.Thing  Analytic_parameters Location Visitor   ∀ hasCountry.Country hasContinent ∃ hasContinent.Thing  Analytic_parameters Location   ∀ hasContinent.Continent hasSource ∃ hasSource.Thing  Analytic_parameters   ∀ hasSource.Source Data Properties Description Logic hasDate ∃ hasDate.DatatypeLiteral  Analytic_parameters Price Product_availability   ∀ hasDate.DatatypedateTimeStamp hasHour ∃ hasHour.DatatypeLiteral  Analytic_parameters   ∀ hasHour.Datatypetime hasNetworkDomain ∃ hasNetworkDomain.DatatypeLiteral  Analytic_parameters   ∀ hasNetworkDomain.Datatypestring hasValue ∃ hasValue.DatatypeLiteral  Analytic_parameters Article_number Price Product_availability For simplicity, we describe here a representative subset of main classes including some of their most interesting object and data properties. These classes are: Analytics_parameters, E-shop, Visitor, Page , and Item . Each class requires a set of properties or conditions in order to be conceptualized. That is, an individual that satisfies those properties is considered to be a member of that class. -Analytics_parameters . Those attributes provided by Google Analytics and Piwik that depend on time. Each analytic parameter has a value ( hasValue in Table 3 ), which corresponds to the data provided by the analytic tool, and a date ( hasDate ), which corresponds to the date when the data was obtained. Subclasses in the ontology (  Analytic_parameters) are, among others: Average_order_value, Average_pages_visited_per_session, Average_session_duration, Average_time_on_site, Bounce_ rate, Conversion_rate, Number_of_transactions, Number_of_landings, Number_of_new_visitors, Number_of_page_views, Revenue_per_ session and Total_revenue . Table 3 shows some representative object and data properties of Analytics_parameters . Each analytic parameter belongs to a data type. For instance, the value of Number_of_transactions is an non-negative integer and the value of Conversion_rate is a float. Data type restrictions are included in the ontology by means of data properties. - E-shop. An e-shop has one or several pages and also an e-shop’s owner. Each e-shop’s owner has an address. Table 4 E-shop group: object and data properties. Object properties Description logic hasVisitor >≡makesVisit > − ∃ hasVisitor.Thing  E-shop   ∀ hasVisitor.Visitor hasNumberOfVisitors ∃ hasNumberOfVisitors.Thing  E-shop Page   ∀ hasNumberOfVisitors.Number_of_visitors hasNumberOfVisits ∃ hasNumberOfVisits.Thing  E-shop Page Visitor   ∀ hasNumberOfVisits.Number_of_visits isOwnerOf ∃ isOwnerOf.Thing  E-shop_owner   ∀ isOwnerOf.E-shop Data properties Description logic hasName ∃ hasName.DatatypeLiteral  Browser Competitor E-shop Goal Item Operating_system Page Product   ∀ hasName.Datatypestring hasURL ∃ hasURL.DatatypeLiteral  Competitor E-shop Page Price   ∀ hasURL.Datatypestring Table 5 Visitor group: object and data properties. Object properties Description logic hasDevice ∃ hasDevice.Thing  Visitor   ∀ hasDevice.Device hasNumberOfVisits ∃ hasNumberOfVisits.Thing  E-shop Page Visitor   ∀ hasNumberOfVisits.Number_of_visits hasCity ∃ hasCity.Thing  Analytic_parameters Location Visitor   ∀ hasCity.City makesVisit hasVisitor > ≡makesVisit > − ∃ makesVisit.Thing  Visitor   ∀ makesVisit.Visit Data properties Description logic hasDaysSinceFirstVisit ∃ hasDaysSinceFirstVisit.DatatypeLiteral  Visitor   ∀ hasDaysSinceFirstVisit.DatatypenegativeInteger hasDaysSinceLastOrder ∃ hasDaysSinceLastOrder.DatatypeLiteral  Visitor   ∀ hasDaysSinceLastOrder.DatatypenegativeInteger hasDaysSinceLastVisit ∃ hasDaysSinceLastVisit.DatatypeLiteral  Visitor   ∀ hasDaysSinceLastVisit.DatatypenegativeInteger ∃ IsReturningVisitor.DatatypeLiteral  Visitor   ∀ IsReturningVisitor.Datatypeboolean Table 6 Visit group: object and data properties. Object properties Description logic hasNavigationStep ∃ hasNavigationStep.Thing  Visit   ∀ hasNavigationStep.Navigation_step hasRefererKeyword ∃ hasRefererKeyword.Thing  Visit   ∀ hasRefererKeyword.Referer_keyword makesVisit hasVisitor > ≡makesVisit > − ∃ makesVisit.Thing  Visitor   ∀ makesVisit.Visit Data properties Description logic hasDuration ∃ hasDuration.DatatypeLiteral  Visit   ∀ hasDuration.Datatypetime hasReturningVisitor ∃ hasReturningVisitor.DatatypeLiteral  Visit   ∀ hasReturningVisitor.Datatypeboolean Attributes for the e-shop are latitude, longitude and time zone. The e-shop’s owner can have competitors, who are eshop’s owners of other e-shops. The analytic parameters of an e-shop are: average_order_value, average_pages_visited_per_ session, average_ session_duration, average_time_on_site, conversion_rate, date_of_last_transaction, number_of_customers, number_ of_failed_transactions, number_of_successful_transactions, number_of_ new_customers, number_of_new_visitors, number_of_sessions_ by_medium, number_of_transactions, number_of_unique_visitors, number_of_units_sold, number_of_visitors, number_of_visits, percentage_of_new_sessions, revenue_per_session, total_revenue , and number_of_returning_visitors . All the analytic parameters related to an e-shop are time dependent. Therefore, they are modeled as classes and related to the e-shop by the corresponding object property. Table 4 shows a subset of properties with classes in the e-shop group as domain. -Visitor and visit . Class visitor has two subclases: customer and New_visitor . A customer is a visitor who makes a purchase. If it is the first purchase of this customer, he/she is a new customer. Customers have an address and name, whereas visitor do not. A visitor visits the e-shop by using a device. The analytic parameters for visitors are: bounced_rate, number_of_visits, and number_of_visited_pages . The analytic parameters for customers are number_of_transactions . Visitors visit pages. Visits are essential to capture the behavior of a visitor when visiting the e-shop. A visit has an entry page and an exit page. It also has a referrer page, which is the way the visitor has accessed the site, i.e. search engine, social network, web-advertisement etc. If the referrer page is a search engine, the keywords used to find the site are also associated with the visit. A visit has a given duration. The attributes of a visit are the times when the entry page and the end page were accessed, duration, back link, whether or not an order was placed and the total goals converted during the visit, number of actions, number of events and number of searches. During a visit, transactions are made. A visit follows a path which has a next page, a previous page and a number. The class Navigation_step is used to model the path that the user follows from the entry page to the exit page. Each navigation step has only one Table 7 Page group: object and data properties. Object properties Description logic hasNumberOfVisits ∃ hasNumberOfVisits.Thing  E-shop Page Visitor   ∀ hasNumberOfVisits.Number_of_visits hasNumberOfVisitors ∃ hasNumberOfVisitors.Thing  E-shop Page   ∀ hasNumberOfVisitors.Number_of_visitors hasTotalRevenue ∃ hasTotalRevenue.Thing  E-shop Page   ∀ hasTotalRevenue.Total_revenue isOnPage ∃ isOnPage.Thing  Item   ∀ isOnPage.Page Data properties Description logic hasName ∃ hasName.DatatypeLiteral  Browser Competitor E-shop Goal Item Operating_system Page Product   ∀ hasName.Datatypestring hasURL ∃ hasURL.DatatypeLiteral  Competitor E-shop Page Price   ∀ hasURL.Datatypestring hasTitle ∃ hasTitle.DatatypeLiteral  Page   ∀ hasTitle.Datatypestring Table 8 Item group: object and data properties. Object properties Description logic hasItem ∃ hasItem.Thing  Page   ∀ hasItem.Item hasPrice ∃ hasPrice.Thing  Item ShareProductData   ∀ hasPrice.Price includes ∃ includes.Thing  Order   ∀ includes.Item isOnPage ∃ isOnPage.Thing  Item   ∀ isOnPage.Page Data properties Description logic hasCategory ∃ hasCategory.DatatypeLiteral  Item   ∀ hasCategory.Datatypestring hasName ∃ hasName.DatatypeLiteral  Browser Competitor E-shop Goal Item Operating_system Page Product   ∀ hasName.Datatypestring hasItemID ∃ hasItemID.DatatypeLiteral  Item   ∀ hasItemID.DatatypenonNegativeInteger hasQuantity ∃ hasQuantity.DatatypeLiteral  Item   ∀ hasQuantity.DatatypenonNegativeInteger attribute number. Tables 5 and 6 show the properties with classes in the visitor and visit group as domain, respectively. -Page. Pages contain items, i.e. product and/or services to be sold. The analytic parameters for Page are: average_order_value, average_time_on_page, bounce_rate, date_ of_last_ visit, number_of_exits, number_of_landings, number_of_new_visitors, number_ of_page_views, number_of_returning_visitors, number_of_sessions_ by_medium (mediums are direct link, social media and search engine), number_of_sessions, number_of_unique_page_views, number_of_unique_visitors, number_of_units_sold, number_of_visitors, number_of_visits, revenue_ per_session_and_total_revenue . Attributes of page are title and URL. A series of representative properties whose domain is page are shown in Table 7 . Interestingly, we can observe in this table that the property hasTotalRevenue is related to the Page , as well as the to the whole e-shop, as this value can be calculated for both classes. -Item. As commented before, an Item is a product or a service which is sold in an e-shop . Specific items of an e-shop are modeled by defining a domain ontology for a specific domain, i.e., travel, books, music, etc. Table 8 contains some representative object and data properties of class item . According to this, Items have a price ( hasPrice ). Prices are valid on a certain date. Therefore, attributes for prices are value, currency and the date for the price validity . The attributes of Items are category and whether or not it has been deleted. Products have a manufacturer . The attributes for products are: name, type, availability on a specific date and article number. The article number can be “ASIN”, “EAN” or “ISBN”. 4.2. Data sources: mapping and querying As we explained in Section 3 , we have focused on three main sources of data coming from different web tracking methods, namely: Google Analytics, Piwik, and specific web scrapping methods in the scope of SME E-Compass project. The process of translating the collected data from different sources to RFD is carried out by means of mapping functions. Each data source has a different set of methods to gather, harmonize, store and provide access to the analytical data. Therefore, a different set of mapping functions is required to parse the information coming from each data source to RDF, according to the ontology. Fig. 3 illustrates an general overview of the mapping process to store data from different sources in a common RDF repository. Each set of mappings is then composed by functions to translate the attributes with their values into their corresponding triple form in RDF. In fact, for most of the attributes, a corresponding mapping function has been developed, involving its corresponding class in the ontology. Nevertheless, as a number of analytic attributes shares a common structure in the ontology, they have been mapped by using generic functions, hence taking advantage of the ontology’s design. Fig. 3. General overview of the mapping process injecting data from different sources into the RDF repository. 4.2.1. Google Analytics Google Analytics 5 is a partially free web analytics service that provides statistics and basic analytical tools for Search Engine Optimization (SEO) and marketing purposes. The service is available to anyone with a Google account, although advanced e-commerce functionalities are only available for restricted users. Google Analytics is geared toward small and medium-sized retail websites. The web tracking procedure in Google Analytics is performed by a “snippet” or digital footprint component, that provides the developer with an API of functions for accessing to each attribute value. This digital footprint is a small piece of JavaScript code that is pasted into the e-shops HTML source code and deployed in the web server where the e-shop is hosted. It activates Google Analytics tracking by inserting the JavaScript ga.js / analytics.js into the page. As illustrated in Fig. 3 , the JavaScript component is then instantiated by our mapping functions by means of a series of java classes to generate RDF triples. Table A.1 in Appendix A contains the set of Google Analytics attributes that are currently tracked by our semantic approach. In this table, each attribute is listed with regards to its corresponding ontology class, data type, and description. This is a representative subset of the whole set of possible attributes (and its combinations) in the Google Analytics’s API specification, that covers all our preliminary requirements for visitor’s behavior and products’ analysis. However, it is worth mentioning that the proposed ontology can be easily extended to consider any of the attributes worked with Google Analytics. 4.2.2. Piwik Piwik 6 is a free and open source web analytics application running on a PHP/MySQL web server. Piwik tracks online visits to one or more websites and displays reports on these visits for analysis. Piwik analytic features include, among others: real-time data updates, free advanced e-commerce analytics features, goal conver5 http://www.google.com/analytics/ 6 http://piwik.org/ sion tracking, event tracking, geolocation, pages transitions views and page overlay. Similarly to Google Analytics, the web tracking procedure in Piwik is also performed by a digital footprint script, that is allocated in the e-shops HTML source code. In the case of Piwik, the analytical data is automatically stored in a relational database (SQL). Therefore, as we have the possibility to access to this relational database, we have developed the mapping functions to directly query the analytic attributes. These attributes are described in Tables A .2–A .4 of Appendix A with regards to their corresponding ontology classes. The obtained data is then translated to RDF according to the ontology by means of specific mapping methods, as shown in Fig. 3 . 4.2.3. Web scrapping methods In the scope of the SME E-Compass project, there exist a series of methods for scraping product and price data from the competitors e-shop websites. This way, a given e-shop’s owner is able to compare their products’ prices with those ones of their direct competitors automatically. This specific functionality provides a REST API service from with we can obtain attributes of competitor’s profile in JSON 7 format, which is a compact and easily readable data format for the purpose of data exchange. Table A.5 contains the competitors attributes that are mapped to RDF in our semantic model (see Fig. 3 ), with regards to the corresponding ontology classes. 4.2.4. RDF repository Finally, an RDF repository is used to integrate the analytic data collected and mapped from the different sources. Therefore, by means of an SPARQL endpoint, it is now possible to query the analytic data unambiguously and independently of the source. As an instance of data access, let us consider an scenario in which the analytic module requires information concerning the 7 http://json.org Fig. 4. Example of SPARQL query that returns disaggregated data attributes, as the ones provided by Piwik, as well as calculated metrics, as those obtained from Google Analytics. Table 9 Two samples of the query result ( Fig. 4 ) of a certain time slot (day 2015-10-23) of a real e-shop. Attribute/metric Visit75688 Visit75692 timestamp 14 :19:44 14 :21:41 visit_total_searches 0 0 visit_total_events 0 0 visit_total_duration 2071 12 visit_total_goal_converted 1 0 total_bounce_rate 52.6066 total_conversion_rate 34.1232 total_number_of_entries 211 total_number_of_new_visitors 145 visits of a given e-shop, in a certain date or period of time. The required information of visits should consist of both: disaggregated data attributes, as the ones provided by Piwik, and calculated metrics, as those obtained from Google Analytics. The SPARQL query represented in Fig. 4 unifies the encoding of such logic, for which a couple of result samples are displayed in Table 9 . In concrete, these results correspond to two consecutive visits to the e-shop with ID < eshop-id > , that were performed at date 2015-10-23. The visit IDs are 75688 and 75692, and they were captured at timestamps 14:19:44 and 14:21:41, respectively. As shown in this table, the visit with a prolonged duration led to one goal conversion (usually a successful sale), whereas the visit with a short duration finished without any conversion, which represents a visitor that leaves the site prematurely. In the case of aggregative attributes, they are calculated for all the visits in the time period of the SPARQL query. Therefore, as shown in the second half of Table 9 , the e-shop registered a bounce rate close to 53% with conversion rate 8 of 34.12%, that corresponds to all visits, bounces and purchases of the queries time period. Another important attribute is the number of new visitors, that for this e-shop and for this date is 145, e. g., 68.72% of total entries. This information could be now used to feed predictive algorithms that help the e-merchant to adopt a given marketing strategy to catch clients. In order to automatize and simplify the accesses to the stored data, our semantic approach includes a specific REST API service with methods that implement predefined SPARQL queries. These methods are used as input of the data mining algorithms as described in the following step of validation. As an additional advantage of this semantic approach, it is possible to connect our RDF repository with other/s external open linked data repository/ies. In this regard, a minimum adaptation has to be done in terms of deciding which class/classes are directly linked from the two repositories with similar semantic meaning. In fact, this is one of the most powerful features when using the semantic structure induced by the ontology. 8 Conversion rate: proportion of visitors converted into paying customers.