— The media industry has undergone a rapid pace of change, leading to an unprecedented explosion of user-generated and platform-generated data. Today’s users interact with digital platforms across a range of channels, including video streaming services, online news portals, social media platforms, and advertising networks. Such interactions are multi-dimensional, dynamic, and influenced by temporal, contextual, and social factors, not linear or uniform. Existing traditional recommendation systems usually utilise limited behavioural signals such as clicks or ratings, which are unable to capture the complex non-linear characteristics of real-world user engagement. In order to overcome these limitations, this paper presents an intelligent user modelling and recommendation framework, called MUMA (Multi-dimensional User Modelling Architecture), specially developed for the changing needs of the media industry.The proposed framework is based on hybrid machine learning techniques and aims to build a spatial-temporal dual-driven user characterisation system. Unlike traditional models that only leverage single-source data, MUMA combines heterogeneous data from multiple sources such as clickstream logs, view duration, social network relations, device meta-data, and eye-movement hotspot signals. While clickstream data reveals explicit interaction patterns, it is viewing duration that offers deeper insights into the intensity of user engagement. Social graph information describes influence relationships and preference diffusion among communities. Eye-movement hotspot data provide implicit behavioural information regarding content attention and cognitive interest. The system combines these disparate signals to build a comprehensive, high-resolution representation of user preferences.MUMA's core is the Dynamic Interest-Aware Network (DIN), which is a dynamic representation of user preferences, not a static one. User interests evolve over time, influenced by changes in context, trending topics, or seasonal effects. The system uses a hybrid LSTM-Transformer architecture with a time decay factor to robustly capture this temporal evolution. The LSTM part learns sequential dependencies and long-term behaviour memory, which enables the model to memorise historical preference information. Meanwhile, the Transformer component leverages self-attention mechanisms to capture short-term interest changes and contextual relevance among recent interactions. The time decay factor weighs recent behaviours more heavily than older ones, reflecting the realistic dynamics of interest. This hybrid structure allows the system to simultaneously capture short-term engagement patterns and long-term preference stability.An additional significant novelty of the MUMA framework is the formulation of a cross-domain migratory learning module based on a Heterogeneous Information Network (HIN). In modern media ecosystems, users interact with multiple modalities of content such as news articles, videos, advertisements and social posts. In general, traditional systems address these domains independently, resulting in disjointed user profiles and suboptimal recommendations. The HIN-based module captures complex relationships between users, items, and contextual entities across different domains. The system leverages graph-based representation learning and transfer mechanisms to transfer learned preferences from one domain to another. For example, a user’s interest in technology news can be leveraged to personalise video recommendations or targeted advertisements. This cross-domain collaboration significantly improves the recommendation coverage, alleviates the data sparsity problem and improves the cold-start performance.Moreover, MUMA combines reinforcement learning and causal inference for a bandit-propensity hybrid recommendation strategy. In recommendation systems, there is a basic trade-off between exploration (introducing new or less-known content) and exploitation (recommending content similar to previously liked items). Too much exploitation can create filter bubbles and user fatigue, while too much exploration can reduce immediate engagement. The system dynamically balances this trade-off by integrating multi-armed bandit algorithms and propensity score-based causal adjustment. Reinforcement learning optimises long-term cumulative rewards, such as user retention and satisfaction, while causal inference mitigates selection bias and guarantees fair content exposure. Such a hybrid decision-making mechanism improves both the short-term click-through rates and the long-term engagement sustainability.
Mrs.C.SHIRISHA et al. (Fri,) studied this question.