Authors: Vinod B. Ingale, Ashish Vankudre, Sagar mali , Dhanaji Jadhav, Pramod Shitole
Abstract: The proliferation of digital platforms has generated vast amounts of event-based temporal data, where user interactions are logged as discrete events in continuous time. Traditional recommendation systems often fail to capture the intricate dynamics of such data, including the exact timing, inter-event gaps, and evolving nature of user preferences. This paper proposes a novel hybrid neural architecture that synergistically integrates Temporal Point Processes (TPPs) with a Self-Attention mechanism to model user temporal behavior for next-item recommendation. Our model, the Temporal Self-Attentive Hawkes Process (TSAHP), leverages the self-attention mechanism to capture complex, long-range dependencies within user interaction sequences, while a neural Hawkes process models the continuous-time intensity of these interactions, inherently accounting for the excitement and decay effects of past events. We evaluate the proposed TSAHP model on two real-world datasets: Amazon Electronics and LastFM. Comparative analysis against state-of-the-art methods, including Time-Aware Matrix Factorization, GRU-based models, and standard Hawkes Process models, demonstrates the superiority of our approach. The TSAHP model achieves significant improvements, with an average increase of 12.5% in Hit Rate @10 and 15.3% in NDCG @10 on the Amazon dataset, and 9.8% in HR@10 and 11.7% in NDCG@10 on the LastFM dataset. The results indicate that explicitly modeling both the semantic context through self-attention and the temporal dynamics via point processes is crucial for accurate and timely recommendations in event-based systems.
