Modeling Temporal Dynamics of User Preferences through Multi-Level Similarity in Recommender Systems

Abstract

Traditional collaborative filtering systems, which rely on user behavioral similarities, often suffer from fundamental limitations the most significant being their neglect of temporal aspects in data analysis. These systems assume that user preferences remain stable over time, assigning equal weight to both old and recent ratings. However, user tastes can change considerably over time. This paper proposes a time-aware movie recommendation approach that addresses these challenges by intelligently integrating both direct and indirect user relationships. Directly similar users are identified based on historical rating data, while indirectly similar users are discovered through dominant opinion pattern mining. A temporal weighting mechanism is applied to dynamically reduce the influence of outdated interactions, aligning recommendations with evolving user preferences. The incorporation of dominant opinion patterns and the analysis of the target user's preferences further enhance the identification of indirectly similar users. Facilitating interconnections and mediation among users helps to mitigate data sparsity issues. Moreover, incorporating time as a key factor enables the system to effectively manage dynamic user behavior and reduce its negative impacts. Scalability concerns are also addressed through the utilization of dominant opinion patterns. Ultimately, by analyzing each target user's preferences, the proposed model delivers more personalized and accurate movie recommendations. Experimental results on the MovieLens dataset demonstrate that the proposed approach significantly reduces the Mean Absolute Error (MAE) and improves prediction accuracy compared to conventional methods.