Authors: Vinod Ingale, Sayli Jadhav, Priyanka Telshinge, Rahin Tamboli, Ashwini Mahind
Abstract: The proliferation of digital platforms has led to an explosion of complex user interaction data, characterized by its sequential nature and rich contextual information. Traditional collaborative filtering (CF) methods often fall short by treating user preferences as static and ignoring the nuanced impact of temporal context and real-world events. This paper proposes a novel recommendation framework, the Temporal-Event-aware Support Vector Machine (TE-SVM), designed to effectively model the dynamic evolution of user preferences by integrating temporal dynamics and event-based contextual signals. The TE-SVM model formulates the recommendation task as a classification problem, where the objective is to find an optimal hyperplane that separates user preferences for items at a given time under specific event conditions. We engineer a comprehensive feature set that captures temporal patterns (e.g., time decay, periodicity) and event embeddings derived from external knowledge sources. A thorough comparative analysis is conducted against established models, including Matrix Factorization (MF), TimeSVD++, and Recurrent Neural Networks (RNN). Experimental results on a large-scale e-commerce dataset demonstrate that the proposed TE-SVM model achieves a significant improvement, with a 12.7% increase in Precision@10 and a 9.8% increase in NDCG@20 compared to the best-performing baseline. The findings underscore the efficacy of SVM in handling high-dimensional, heterogeneous feature spaces for temporal and event-aware recommendation tasks, providing a robust and interpretable alternative to deep learning-centric approaches.
