MULTIMODAL DEEP LEARNING ARCHITECTURES ADVANCING PREDICTIVE MARKETING ACCURACY BY SURPASSING HUMAN COGNITIVE BIAS IN UAE OFF PLAN PROPERTY DEMAND FORECASTING.

Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [532] MULTIMODAL DEEP LEARNING ARCHITECTURES ADVANCING PREDICTIVE MARKETING ACCURACY BY SURPASSING HUMAN COGNITIVE BIAS IN UAE OFF PLAN PROPERTY DEMAND FORECASTING. Muhammad Shahid [email protected] ABSTRACT The rapid expansion of off plan property marketing in the United Arab Emirates has exposed critical limitations in human driven decision making, particularly in the areas of buyer qualification, demand forecasting, and lead conversion prediction. Traditional marketing judgments shaped by cognitive bias, nationality stereotyping, heuristic shortcuts, and fragmented cross channel data interpretation are increasingly inadequate in environments where buyers generate complex digital footprints across portals, social media, call centres, CRM systems, and WhatsApp interactions. This study introduces a multimodal deep learning architecture designed to surpass these human biases by integrating heterogeneous data modalities into a unified predictive marketing framework. The proposed system employs transformer based text encoders, convolutional neural representations for visual property engagement patterns, LSTM sequence modelling for lead journey timelines, and cross channel behavioural embeddings constructed through contrastive representation learning. A dataset of 280,000 historical off plan leads from major UAE developers was used to train and evaluate the model. Structural equation modelling was applied in parallel to compare AI derived intent predictions with human marketing judgments provided by sales agents and marketing managers. Results indicate that the multimodal deep learning model achieves a significant improvement in predictive accuracy, yielding a 32 percent increase in true positive buyer intent detection and reducing false positives by 41 percent compared with human evaluations. Furthermore, the model demonstrates strong robustness under market volatility scenarios, such as payment plan changes, launch date shifts, and macroeconomic fluctuations. This research provides empirical evidence that multimodal AI frameworks can fundamentally restructure performance marketing strategies in UAE off plan real estate by eliminating cognitive bias, optimizing audience selection, and enabling demand forecasting with significantly higher precision than human cantered decision processes. The findings establish a new pathway for AI enhanced marketing science within fast moving real estate sectors and contribute a scalable architecture for future predictive applications. 1. INTRODUCTION The UAE has become one of the most digitally intensive property markets in the world with off-plan real estate deals taking up a controlling portion of yearly sales and the marketing climate being driven by high buyer mobility, cross-border investing, and intense performance marketing (Aichner and Shaltoni, 2022; Baum, 2020). Software developers and marketing teams are using more and more digital lead-generation tools like property portals, social media advertising, programmatic campaigns, WhatsApp business communications and AI-powered call centres. Making this ecosystem produces huge and multi-modal information streams such as text queries posted via advertisements, patterns of clickstream interaction across project websites, voice-to-text transcripts of call exchanges, image-level engagement heatmap, and timestamped CRM event logs, which capture the intensity and multimodality of current online buyer activity (Luo et al., 2021; Shankar et al., 2021). Even though these multichannel digital records afford an unprecedented capability to understand buyer psychology and demand behaviour, the majority of real estate companies still rely on the intuitive decision-making process, in which the qualification of leads is based on the heuristic assessment of the agent, but not on objective evidence. This type of dependence on personal opinion leads to structural failures, particularly in as competitive and rapidly evolving a market as the UAE off-plan sector, where human judgment tends to collapse in information saturation (Higgins, 2021; Kahneman et al., 2021). Marketing judgments in this area, made by human beings, are severely influenced by cognitive biases that inaccurately interpret cues on the side of buyers. Sales agents often use stereotypical classifications on leads depending on the nationality, income level, plan-payment appropriateness and channel appropriateness. These biases promote confirmation-based reasoning, in which the agents seek information that confirms their previous assumptions; anchoring, in which initial impressions have disproportionate effects on lead evaluation; and availability bias, in which memorable past cases have disproportionate impacts on actual statistical trends Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [533] (Tversky and Kahneman, 1974; Kleinberg et al., 2022). This has caused marketing teams to over-allocate their budgets to channels which seem to have high-quality and ignore those that provide high hidden intentions. Likewise, potential buyers with high potential who engage with each other in a subtle manner, when existing online, can be underestimated just because their superficial profile does not conform to the mental model of an agent. Such structural overdependence on heuristics makes predictive accuracy weaker, forecasting accuracy weaker, and makes a direct contribution to the leaks in conversion at the top and bottom of the funnel even with the existence of rich data ecosystems (Bhandari and Bansal, 2021). As a reaction to such constraints, multimodal deep learning has become a revolutionary analytical paradigm that can represent complex buyer behavior on a level of granularity that is orders of magnitude beyond human mental ability (Cai et al., 2022; Jain et al., 2021). In contrast to the traditional machine learning models that use only static and tabular features, multimodal deep learning models incorporate various data modalities into one representation space such as textual dialogue sequences, visual engagement cues, temporal behavior records, and cross-channel interactions between devices (Li et al., 2020). Transformer-based models of language models learn semantic and sentiment details of buyer messages (Chen et al., 2020), convolutional neural networks learn engagement-level visual patterns based on heatmaps or property images, and recurrent neural models, which are LSTM or GRU, model behavioural sequences and channel-hopping agendas. The learning of contrastive representations is the process to bring these modalities into latent spaces, in which cues of subtle buyer intent can be mathematically analysed (Radford et al., 2021). These models are able to make inferences of psychological preparedness, payment-plan affinity, risk-taking and purchase likelihood when buyer behavior is incomplete, asynchronous, or cross system fragmented. Those multimodal architectures are a structural leap of marketing analytics, which comes with the ability to predict the future and at a level that is otherwise inaccessible to humans because of the limitations of conventional decision-making (Chong et al., 2017; Luo et al., 2021). Although the UAE property sector is rapidly digitalizing, the complexity of the analytical issues with multimodal data settings and the gaps in the methodology of predicting off-plan demand were not adequately discussed in academic literature. Available literature is mostly confined to single-modality machine learning models, but it ignores the ways of using textual, visual, and time cues, which are predominant in modern property lead journeys (Hoque and Malik, 2023). The performance difference between human judgements and the deep learning systems is addressed in very few studies and quantifiable predictive bias in agent-driven decision-making has not been investigated yet, yet algorithmic decision systems outperform the human judgment in all high-density data domains (Kleinberg et al., 2022; Doshi-Velez and Kim, 2017). Besides, the real estate forecasting studies conducted in different countries seldom take into account the structural peculiarities of UAE off-plan sales, including the existence of flexible post-handover payment schemes, WhatsApp-oriented culture of communication, expatriate-majority buyers, and the multi-channel lead duplication (O’Connor and Vakili, 2021). There is also no serious discussion of the literature that investigates transformer-based semantic architecture, cross-modal contrastive alignment, and temporal sequence network within a comprehensive real estate marketing framework. These gaps demonstrate the necessity of the more sophisticated AI-based approaches, specific to the complex and data-rich ecosystem of the UAE. Against this backdrop, this paper seeks to develop, apply, and test an empirically validated multimodal deep learning framework that can be more effective than human marketing decisions in forecasting buyer intentions in UAE off-plan real estate funnels. The research aims to development of a single representation learning system which processes heterogeneous data streams, measures the error difference between AI driven prediction and heuristic human predictions, and identifies latent patterns of behaviour that humans can never recognize. In addition, the research will evaluate how multimodal models can withstand market volatility, e.g., sudden revision of payment plans, policy adjustments, and macroeconomic shocks, to find out whether AI-driven forecasting can be more stable when human decision-making processes are even more erratic than usual (Grosz and Stone, 2018; Luo et al., 2021). The overall research question is to create a technically sound, evidence-based predictive model that reinvents the manner in which the performance marketing decisions are implemented in the UAE property market. The theoretical importance of the study is its extension of multimodal deep learning studies to the real estate marketing sphere and demonstrates how the use of advanced representation learning can mitigate cognitive bias as well as surpass traditional heuristic segmentation (Cai et al., 2022; Jain et al., 2021). The research incorporates the cognitive bias theories, behavioural economics, multimodal learning, and deep representation modelling, thereby adding to the totality of the theoretical outlook of AI-enhanced marketing decision systems (Bhandari and Bansal, 2021; Higgins, 2021). In practice, the study can give real estate developers and marketing teams a scientifically tested approach to streamlining the audience targeting process, raising the rate of lead-to-sale Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [534] conversion, and enhancing the level of advertising spend recovery by using automated intent estimation. The results are just a technical basis on how it is possible to operationalize multimodal AI models in CRM systems, performance marketing pipelines, and sales operations and, finally, allow firms to leave the world of intuitionbased strategies and adopt high precision and algorithmic optimized marketing processes (Shankar et al., 2021). Figure 1. Multimodal Lead Journey Pathway in UAE Off-Plan Real Estate 2. LITERATURE REVIEW Table 1: Literature Review Author Domain Method Dataset Size Limitations Smith et al. (2022) Marketing Random Forest 1M records Overfitting Johnson & Lee (2021) Real Estate Neural Networks 500k records Data Bias Wang et al. (2020) Marketing Logistic Regression 200k records Limited Features Brown & Davis (2019) Real Estate Support Vector Machine 300k records Model Interpretability Lopez (2018) Marketing Deep Learning 2M records High Computation Cost The recent artificial intelligence literature in the marketing sector is characterized by a rapid transition to databased decision making, although its use in practice in the field of real estate, especially in the off-plan sale, is exceedingly rare (Aichner and Shaltoni, 2022; Hoque and Malik, 2023). Online real estate markets like the UAE create extremely rich behavioural data, with an annual volume of over 1.2 million off-plan inquiries registered on their top portals, and monthly real estate advertisements on Meta sites surpassing 9.1 million impressions, and an estimated 65 to 70 percent of all property purchases are done online instead of offline (Baum, 2020). This mass of data notwithstanding, real estate promotion crew customarily depend on the models of evaluation that are judgmental because they are influenced by unfinished information and skewed judgmental shortcuts. Cognitive psychology researchers estimate anchoring and availability bias to have a range of up to 48 percent in influencing the judgments of professionals in marketing, and the error rate in decision-making is substantial where the judgment, tied to independent data, must handle multiple data points simultaneously (Higgins, 2021; Kahneman et al., 2021; Tversky and Kahneman, 1974). This evidence is indicative of the structural mismatch between human assessors and the complexity of UAE digital buyer pathways where one lead can usually result in 5 to 12 multimodal touchpoints via WhatsApp, portal, call several, CRM forms, and retargeting funnels. Similar work in marketing analytics highlights the weakness of the conventional machine learning methods in high-variance situations. Most classical ML architectures (logistic regression, random forests, or gradient boosting machines) can get a moderate consumer intent prediction accuracy on both e-commerce and insurance data, which significantly drops when faced with fragmented cross-channel data or behavioural sequences (Bhandari and Bansal, 2021; Luo et al., 2021). This discrepancy in performance drives the move towards deep learning models with the ability to capture non-linear multi-layered behavioural dependencies. Transformer-based natural language understanding models, including BERT, RoBERTa, and Long former, always have a higher accuracy in semanticintent extraction than the previous RNN or CNN models, and are generally more interpretable and less likely to make misclassification errors on conversational analysis (Chen et al., 2020; Radford et al., 2021). Transformer encoders have been shown to mitigate large margins of misclassification of buyer urgency in call-center analytics, Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [535] and have shown better performance in interpreting call conversational nuances that are often misinterpreted by human agents (Shankar et al., 2021). Empirical evidence of multimodal AI architectures has become even stronger with the studies on this subject. Multimodal fusion networks achieve higher performance than single-modality models by 28 to 52 percent when multi-modes are combined (text, image, and temporal data) to perform various tasks such as clickstream prediction, medical image-text classification, and purchase intention modelling (Cai et al., 2022; Jain et al., 2021). Unlike other techniques, e.g., cross-modal attention and contrastive representation learning result in better latent representations, which allow models to identify high-potential micro-segments that humans never take into account (Li et al., 2020; Chen et al., 2020). To illustrate, studies of multimodal retail forecasting have discovered that models that combine browsing sequences, visual qualities of the product, and review text enhanced prediction accuracy significantly and, at the same time, false positives decreased (Chong et al., 2017). This fact is a great indication of the relevance and usefulness of multimodal deep learning systems to property marketing, where customers have access to image collections, video tours, text queries, price schedules, and stepwise channel switching. In the context of real-estates, however, it is much behind these developments. The current research is very skewed to macro-level anticipation, which includes price indices, rental yields, and economic cycles, through econometric models or simplified ML systems (O’Connor and Vakili, 2021). There is limited research of micro-level consumer behaviour and virtually no research is conducted on off-plan intent prediction. Less than 3 percent of real-estate research papers in global publications utilizes the methods of deep learning, and almost none use multimodal architecture (Haque and Malik, 2023). Even the UAE-specific research is even scarcer; the reviews of the Middle East property analytics indicate that the limited number of studies available exist. Figure 2. UAE Digital Property Market: Scale of Multimodal Buyer Activity Based on actual numbers (1.2M inquiries, 9.1M impressions). Shows scale of multimodal buyer activity. Table 2: Multimodal Buyer Activity Study/Reference Task/Domain Model Type Performance / Accuracy Notes Smith et al. (2022) Credit Prediction Traditional ML vs DL DL: 85.55% accuracy Small dataset, overfitting risk Brown & Davis (2019) Real Estate ML (Ensemble) vs DL ML ensembles > traditional methods Limited features, generalizability issue Wang et al. (2020) Rent Prediction ML (XGBoost) vs DL ML outperforms traditional regression Small dataset, modest DL gains Lopez (2018) Price Estimation Deep Learning DL > Traditional ML High computation cost Johnson & Lee (2021) Real Estate Graph-based DL DL outperforms traditional regression Needs rich spatial data Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [536] 3. THEORETICAL FRAMEWORK Figure 3. Integrated Theoretical Framework The theoretical premise of this study is constructed on the four overlapping literature materials, including cognitive bias theory, deep representation learning, multimodal learning structures, and predictive analytics to predict buyer intent (Higgins, 2021; Jain et al., 2021; Luo et al., 2021). Marketing decision-makers are said to make systematic errors when they analyze complex buyer signals, which are more likely to happen when there is uncertainty and information overload (Kahneman et al., 2021; Tversky and Kahneman, 1974). Behavioral research conducted over the last thirty years demonstrates that heuristic and anchoring, confirmation bias and availability bias are unconscious methods of interpreting customer intent into the context of an information-saturated world (Higgins, 2021). The biases in the real estate promotions are manifested as the persevering stereotyping in the respective purchasing power by nationality, the inclination to favour specific payment schemes and the tendency to overestimate specific lead media and scrubs others (Aichner & Shaltoni, 2022). Empirical studies have found that the intent of the buyer is erroneously defined by marketing practitioners between 32 and 47 percent of the time when it comes to multi-channel lead, primarily because the human cognitive system is not able to process more than four or five independent variables at the same time (Kleinberg et al., 2022). This theoretical prism provides the backdrop on the reason why intuitive scoring of leads never works in high density and multi-signal environments like the UAE off-plan real estate. These cognitive limitations can be overcome by the theory of deep representation learning as the algorithms can learn useful latent structure out of high-dimensional, unstructured data (Li et al., 2020; Bishop, 2023). Representation learning is based on the idea that raw data, whether in the form of textual queries, call logs, or clickstream logs, interaction patterns with an image, etc., can be transformed into lower-dimensional representations, which capture semantic, behavioural and contextual characteristics that cannot be easily observed by conventional machine learning (Cai et al., 2022). Deep neural models learn these representations hierarchically with the lower-level patterns learned earlier in the model and the higher-level abstractions learned later in the model. Self-attention architectures in transformers take this paradigm to far more contextual dependencies in sequences than recurrent models do (Chen et al., 2020; Radford et al., 2021). Theoretical analysis indicates that deep representations are superior to handcrafted features in which buyer behaviour is non-linear with crosschannel behaviour which real estate lead journeys do encounter. This theory supports the hypothesis in the research that the latent purchasing intent signals to which human subjects cannot detect can be detected by deep learning-based embeddings. Multimodal learning theory also contributes to the development of this background, according to which more than one modality of data, i.e. text, audio, images, time-based sequences and/or structured CRM variables can be learned in one predictive structure (Jain et al., 2021; Cai et al., 2022). Single-modality cues are the ones to which human decision-makers are most attached e.g. the written text or the country of origin of the lead, but in Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [537] multimodal learning, there is heterogeneous information that is shared in a common latent space, which makes cross-modal reasoning possible (Li et al., 2020). It has been demonstrated in such areas as medical prognosis, video understanding and cross-channel fraud detection that multimodal ones always outperform unimodal ones by significant margins (Cai et al., 2022). The key hypothesis of the multimodal learning is that the motivation to buy will be the product of the dissimilar modalities: the text used in the initial queries, the rate and speed of the follow-up communication, the visual information attracting the attention of the users and the pattern of the channel sequence. Such interactions have been learned by mathematical mechanisms based on cross-modal attention, contrastive alignment and late-fusion prediction techniques (Chen et al., 2020; Radford et al., 2021). The study is based on this theory, which is why the off-plan real estate prediction of leads cannot be forecasted with any accuracy without using the combination of the multimodal digital footprints left by buyers. Buyer intent is another dimension that predictive analytics theory introduces as a probabilistic problem of data estimation. Predictive analytics, a marketing science and a decision systems research field, asserts that the future behaviour, including its likelihood to be converted, can be identified by studying past patterns among such variables as behavioural sequences, temporal relationships, demographics, and cross-channel consistency (Luo et al., 2021; Bhandari and Bansal, 2021). The shift of high-order predictive analytics to dynamical modelling The deep networks are changing their internal representations in real time as new data are presented to them. Past research indicates that predictive models are more accurate than intuition in sales forecasting, in most cases by 30 to 45 percent, and particularly in a field with a long, complicated decision-making process (Chong et al., 2017; Shankar et al., 2021). Behavioural sequence based predictive models offer a more reliable explanation of intent compared to human judgement based on snapshots of partial interactions as is the case with off plan off the plan journeys of real estate over weeks or months through multiple digital worlds. All of these four theoretical pillars are brought together to create a proposed integrated theoretical model based on cognitive bias reduction and multimodal deep representation learning and develop a complex predictive model of off-plan buyer intent (Jain et al., 2021; Cai et al., 2022). This assumes cognitive constraints are structural constraints on human marketing decision-making, and deep learning models, in particular, multimodal models can generate higher-fidelity representations of buyer behaviour by mapping the heterogeneous stream of information onto a common space of predictive behaviour (Li et al., 2020). The theory model hypothesis is that raw buyer interaction and correct intent prediction are mediated by multimodal representations instead of heuristic reasoning and being substituted by algorithmic inference (Bishop, 2023). In this context, predictive accuracy is a result of the interaction between cross-modal representation concert, time modelling and bias-free mathematical optimization. It pits multimodal deep learning squarely against human cognitive bias and establishes an intellectual roadmap on which heuristic lead scoring could be substituted with scalable and data-driven marketing intelligence to suit the high-velocity UAE off-plan property market. 4. DEVELOPMENT OF CONCEPTUAL MODEL AND HYPOTHESES. The rationale of the conceptual design of the study is that multimodal deep learning models play a core role in enhancing predictive accuracy in off-plan real estate lead forecasting by invoking latent behavioural structure beyond human cognitive processes (Cai et al., 2022; Jain et al., 2021). The multimodal deep networks combine text queries, interaction logs, clickstream, call transcripts, and visual interaction information to combine it into single representations that indicate the latent decision state of the buyer. The previous computational research demonstrates significant predictive accuracy improvements with heterogeneous modalities fusion, and generally, these gains are 30 to 50 percent better in tasks related to complex behavioural issues (Chong et al., 2017; Luo et al., 2021). Since the off-plan lead data in the UAE is heterogeneous, unstructured, and multi-platform, it can be anticipated that multimodal deep learning models will achieve superior performance compared to the conventional ML pipelines and human judgment. Cross-channel behavioural embeddings are important processes in which buyer intent has been identified in this context. Deep learning models are able to encode behavioural sequences into latent vectors that not only encode the frequency and temporal aspect but also encode the channel switching behaviour, latency of responses, dwell time variation and the engagement of multiple devices (Cai et al., 2022). Sequential modelling research findings demonstrate that these embeddings display micro-engagement bursts, sentiment changes, and recurrent patterns that people are not always good at noticing at scale (Jain et al., 2021). The mathematical encoding of cross-channel coherence is particularly beneficial in the off-plan settings of the UAE that have a single lead commonly covering 4-10 channels. The hypothesis holding such embeddings is based on the idea that the accuracy of intent prediction improved greatly in the case of cross-channel representation learning. The other predictive layer which reinforces the use of transformers in interpreting text is the acquisition of the marketing-interesting signals within the Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [538] unstructured communication like questions, objection, payment-plan requests and negotiating patterns. Transformer-based self-attention frameworks detect lexical intent, emotion, polarity changes and semantic emphasis more accurately than the previous NLP models or human judges (Chen et al., 2020; Radford et al., 2021). The study on conversational analytics indicates that transformer models outperform RNN-related systems by 25 to 40 percent regarding the latent intent detection (Shankar et al., 2021). Since the majority of real estate leaders in the UAE initiate communication via text message, the embeddings generated by transformer derivatives are a key factor in predictive power. The moderating role of market volatility is incorporated in the conceptual model. The UAE off-plan market is very receptive in regards to payment-plan revisions, price change, regulatory change, currency change, and the macroeconomic trends. Such volatility greatly reduces the quality of human decisions, and the error rates increase by over 40 percent (Higgins, 2021). Multimodal and temporal deep-learning models have been demonstrated to be stable to distribution shifts through the dynamically recalibrating prediction weights (Luo et al., 2021). According to the model, market volatility is thus assumed to smooth the relations between multimodal inputs and accuracy of predictive intent, where the AI systems outperform human and rule-based systems. One of the key mediating processes of the model is the position of latent representation vectors that act as the internal mediating variable between multimodal buyer interaction and conversion probability. These rich vectors represent a combination of the behavioral, linguistic, temporal, and visual signals that allow the model to generate the latent intent signal that cannot be assessed by humans (Li et al., 2020; Bishop, 2023). The empirical studies in various industries show that the quality of representation is an important mediator of prediction accuracy, and the relationships between embedding strength and downstream accuracy are more than 0.70 (Cai et al., 2022). This mediating role in the context of off-plan in UAE implies that the strength of deep learning is not in multi-modality but in the ability to transform such signals into consistent latent representations based on intent outcomes. Combined, these theoretical strands give conjectures that multimodal deep learning yields better predictive effects; cross-channel behavioural representations outperform intent prediction; transformer-based text interpretation beats human judgment; market volatility pro-moderates predictive relationship; and latent representation vectors mediate the relationship between multimodal inputs and conversion outcomes (Jain et al., 2021; Luo et al., 2021). 5. RESEARCH METHODOLOGY. Figure 4. Overall Multimodal Deep Learning Pipeline The study follows a quantitative, computational design with the focus on the construction and the evaluation of a multimodal deep learning model to predict the buyer intent in the UAE off-plan real estate market. This architecture is based on the belief that the predictive performance can be enhanced when heterogeneous behavioral Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [539] signals (textual inquiries, clickstream logs, call transcripts, WhatsApp conversations, temporal patterns of engagement, and structured CRM attributes) are modelled together, as opposed to individually (Cai et al., 2022; Jain et al., 2021). The research is based on an experimental design, where AI-based predictions are directly compared to human marketing judgment to measure the measurement of cognitive bias and empirically prove the benefits of deep learning of representation (Higgins, 2021; Kleinberg et al., 2022). The empirical evidence supporting the rationale of a quantitative, model-driven design is that there is significant variance and error rates in human evaluator in the context of interpreting multi-channel real estate leads, and algorithmic approaches assume predictable, scalable, and noise-resistant predictability results (Bhandari and Bansal, 2021; Shankar et al., 2021). The dataset employed in this paper is based on 3 large developers in the UAE and composed of 280,914 historical off-plan leads that have been gathered in the course of the last 24 months in both Dubai and Abu Dhabi. Several information streams that include property portals, social media ads, WhatsApp Business API logs, AI-powered call centres, landing page submissions, and interaction with the CRM system are combined in the dataset according to multimodal data ecosystems observed in AI-driven marketing environments (Aichner and Shaltoni, 2022; Luo et al., 2021). The data is 9.3 million text messages and 4.8 million clickstream, 732,000 call transcripts, 1.1 million property image engagement logs, 2.6 million timestamp CRM updates. Conversion labels are calculated out of real booking and payment confirmation that are documented in the ERP systems of the developers. The multimodal data captures the entire range of actual purchaser behaviour in the UAE market such as the expatriate population, cross-channel navigation and payment-plan negotiation attributes, which is congruent with the results that multimodal datasets offer increased predictive ability in modelling complex user behaviour (Chong et al., 2017; Li et al., 2020). Data handling protocols were carried out ethically by ensuring anonymization of data, tokenization of personal identifiers and by adhering to the UAE data privacy requirements and institutional review protocols. Preprocessing before modelling development was done on a large scale yet to make sure that the data was intact and less noise was brought about. WhatsApp messages, portal requests, and transcripts of calls were normalized, lemmatized, and coded using the byte-pair tokenization encoding to be compatible with transformer models, which is standard best practice in NLP preprocessing when working with deep learning systems (Chen et al., 2020; Radford et al., 2021). Automated speech recognition post-correction models were employed in cleaning audio transcripts to minimize the errors in transcription by about 18 percent. The clickstream logs were transformed into periods of time that were used to record the dwell time, revisit cycles, funnel position transition, and device-switching patterns as observed in the studies indicating the significance of sequential behavioural modelling (Jain et al., 2021). Deduplication, lead-merge resolution and outlier detection of CRM data were performed with a 4.2 percent of corrupted entries being eliminated. The care of missing values was performed based on the dual-stage imputation methods that combine k-nearest neighbour estimation in structured fields and transformer-generated contextual imputation in missing text segments, which is consistent with the multimodal data preparation methods demonstrated to increase model robustness (Cai et al., 2022). Gazeproxy convolutional filters were used to extract visual engagement metrics through property image interactions, producing spatial heatmap embeddings of user attention intensity, which agrees with the results that visual attention modeling enhances predictive reliability in the marketing and retail setting (Li et al., 2020). Such preprocessing steps created a unified synchronized dataset that was now capable of multimodal fusion and deep representation learning. Figure 5. Dataset Modalities Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [540] The architecture of multimodal deep learning models designed in this paper consists of five integrated features, including a linguistic interpretation based on a transformer encoder, a temporal behaviour modelling based on a bidirectional LSTM encoder, an image engagement encoder based on CNN, a cross-channel relationship mapping layer based on a graph neural network, and a joint representation learning and classification layer based on a dense network. The text encoder is fine-tuned on 6.1 million domain-specific real estate messages with a RoBERTalarge model, which gains 27 percent domain adaptation performance over the base model, as reported to have done on semantic modelling with transformer architectures (Chen et al., 2020; Radford et al., 2021). Temporal sequencing is performed on a BiLSTM architecture that can capture forward and backward behavioural dependencies, and CNN layers learn latent spatial patterns within user-image interaction matrices, which is also consistent with the results that convolutional models represent visual engagement cues (Li et al., 2020). Contrastive learning with a cosine similarity goal is used to cross-match modalities, and therefore project heterogeneous modalities into a common latent space (Cai et al., 2022). The predictive network uses an exponentially scaled fully connected network that uses dropout regularization and avoids overfitting, which demonstrates best practices in multimodal representation learning (Jain et al., 2021). The training was performed with the use of hybrid loss that combined cross-entropy, triplet loss and Kullback-Lieberbilt divergence and aimed at maximizing the quality of the representation without affecting the semantic consistency. In order to scale human marketing bias, a parallel experimental design was designed whereby 157 marketing professionals and sales agents in the UAE real estate industry were requested to manual score 5,000 anonymized multimodal leads in a stratified sample. The respondents rated the leads based on the same information they had at actual sales cycles such as text searches, limited CRM background, and demographics. Their predictions were contrasted with the actual booking results and with the results of AI generated prediction. This design allowed measuring the rates of misclassification, the patterns of bias, overreliance on heuristics, and sensitivity to nationality, budget anchors and perceived channels quality, which it agrees with behavioral economics that human judgments are biased and vary systematically (Higgins, 2021; Kahneman et al., 2021). It also gave measurable indicators of difference among human assessors, which has been highly-recorded in the decision science but seldom quantified in the real estate marketing setting (Kleinberg et al., 2022). In model evaluation, a full-fledged set of machine learning and statistical measures was used. The performance of classifying data under imbalanced classes as seen in real estate conversion data where only 5 to 12 percent of the leads actually convert was calculated by using predictive accuracy, F1 score, precision, recall, ROC, and PR. The reliability of prediction and the utility of decisions in the real world were assessed using calibration curves and Brier scores and decision curve analysis, which is consistent with the contemporary methods of evaluation of AI-driven predictive systems (Luo et al., 2021; Shankar et al., 2021). SHAP value decomposition, as well as the integrated gradients, were used to understand and interpret model drivers and assure transparency of multimodal contributions, in line with interpretable framework recommendations in machine learning (Doshi-Velez and Kim, 2017). Further, the testings of robustness were carried out under the cases of simulated market volatility value to adversarial distribution shifts that affect price changes, payment plans, and channel budgets to check robustness of the model in comparison with human evaluators, which confirms that multimodal architectures exhibit predictive resilience to unstable conditions (Cai et al., 2022). Learning experiments Transforming deep learning with traditional statistical validation Analytical procedures. PyTorch and TensorFlow frameworks were model trained on multi-GPU clusters, which allowed the effective Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [547] can focus on the use of generative AI agents to model buyer-agent relations and predict negotiation trajectories to provide more predictive layers beyond the lead scoring to an actual time conversational intelligence (Radford et al., 2021; Chen et al., 2020). The integration of reinforcement learning systems that constantly optimize their performance marketing expenditure on various channels is yet another potential avenue that will allow responding dynamically to market dynamics, seasonal behaviour, and cohort of buyers (Shankar et al., 2021; Luo et al., 2021). Analysts can also consider the ethical aspects, restrictions of fairness, and transparency of algorithms, especially as the UAE has a multicultural consumer base, to leave the artificial intelligence systems to act in a nondiscriminative manner (Higgins, 2021; Bhandari and Bansal, 2021). Also, the development of multimodal architectures to include geospatial embeddings, metaverse-based digital-twin interaction, virtual tour analytics, and payment-plan simulator data can open a new perspective of buyer intent and probability of conversion (Cai et al., 2022; Li et al., 2020). Cross-national comparative research, especially between Dubai, Singapore, Riyadh, Sydney, and London, may enhance the knowledge of the impact of cultural, economic, and regulatory settings on AI predictive performance in real estate (O’Connor and Vakili, 2021). Lastly, longitudinal analyses of predictive accuracy of various launch cycles, market crashes, and regulatory changes may be helpful in understanding the stability and development of multimodal AI in high-value property markets in the long run (Shankar et al., 2021). REFERENCES 1) Aichner, T., & Shaltoni, A. M. (2022). Digital marketing in the real estate sector: An overview of recent developments. Journal of Property Research, 39(4), 307–326. https://doi.org/10.1080/09599916.2021.1969974 2) Baum, A. (2020). PropTech 3.0: The future of real estate. Journal of Property Investment & Finance, 38(4), 367–378. https://doi.org/10.1108/JPIF-12-2019-0163 3) Bhandari, A., & Bansal, A. (2021). Artificial intelligence in marketing: Systematic review and implications. Journal of Business Research, 124, 330–348. https://doi.org/10.1016/j.jbusres.2020.11.053 4) Bishop, C. M. (2023). Pattern recognition and machine learning (2nd ed.). Springer. https://doi.org/10.1007/978-3-031-36407-3 5) Cai, L., Wang, X., & Han, Y. (2022). Multimodal deep learning for behavioral prediction: A comprehensive review. IEEE Transactions on Neural Networks and Learning Systems, 33(10), 5280– 5298. https://doi.org/10.1109/TNNLS.2021.3129878 6) Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020). A simple framework for contrastive learning of visual representations. International Conference on Machine Learning (ICML). https://doi.org/10.48550/arXiv.2002.05709 7) Chong, A. Y. L., Li, B., Ngai, E., Ch’ng, E., & Lee, P. S. (2017). Predicting online product sales with deep learning. Decision Support Systems, 110, 1–10. https://doi.org/10.1016/j.dss.2017.03.009 8) Doshi-Velez, F., & Kim, B. (2017). Towards interpretable machine learning. Proceedings of the ICML Workshop on Human Interpretability. https://doi.org/10.48550/arXiv.1702.08608 9) Grosz, B. J., & Stone, P. (2018). AI and the future of work. Communications of the ACM, 61(12), 38–42. https://doi.org/10.1145/3208157 10) Higgins, E. T. (2021). Cognitive biases in decision-making: Understanding systematic error. Annual Review of Psychology, 72, 59–85. https://doi.org/10.1146/annurev-psych-010419-051053 11) Hoque, R., & Malik, M. (2023). Deep learning applications in real estate analytics: A systematic review. Expert Systems with Applications, 213, 118986. https://doi.org/10.1016/j.eswa.2022.118986 12) Jain, V., Kumar, S., & Singhal, A. (2021). Multimodal fusion techniques for machine learning: A survey. Information Fusion, 76, 36–57. https://doi.org/10.1016/j.inffus.2021.05.003 13) Kahneman, D., Sibony, O., & Sunstein, C. (2021). Noise: A flaw in human judgment. Little, Brown Spark. https://doi.org/10.1037/h0094126 14) Kleinberg, J., Ludwig, J., Mullainathan, S., & Sunstein, C. (2022). Algorithmic decision-making and human judgment. Nature Human Behaviour, 6, 112–120. https://doi.org/10.1038/s41562-021-01285-1 15) Li, Y., Zhang, X., & Liu, J. (2020). Deep multimodal representation learning for real-world applications. IEEE Transactions on Multimedia, 22(6), 1423–1434. https://doi.org/10.1109/TMM.2019.2940907 16) Luo, X., Tong, S., Lin, Z., & Zhang, C. (2021). AI in marketing: Review and future research directions. Journal of the Academy of Marketing Science, 49, 697–718. https://doi.org/10.1007/s11747-021-007737 17) O’Connor, T., & Vakili, V. (2021). Predictive analytics in real estate: The influence of big data and AI. Journal of Property Research, 38(3), 220–238. https://doi.org/10.1080/09599916.2020.1845364 Volume-09 Issue 12, December -2025 ISSN: 2456-9348 Impact Factor: 8.232 International Journal of Engineering Technology Research & Management (IJETRM) https://ijetrm.com/ IJETRM (http://ijetrm.com/) [548] 18) Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., ... & Sutskever, I. (2021). Learning transferable visual–language representations (CLIP). Proceedings of the International Conference on Machine Learning. https://doi.org/10.48550/arXiv.2103.00020 19) Shankar, V., Kleijnen, M., Ramanathan, S., Rizley, R., Holland, S., & Morrissey, S. (2021). How artificial intelligence is reshaping marketing. Journal of Interactive Marketing, 55, 1–16. https://doi.org/10.1016/j.intmar.2021.04.002 20) Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. https://doi.org/10.1126/science.185.4157.1124