Self-Attentive Sequential Recommendation Models Enriched with More Features
Trong Dang Huu Ho, Thi Thanh Sang Nguyen · 2024
Recommender systems have become essential for alleviating information overload by providing personalized suggestions, with traditional collaborative filtering methods based on matrix factorization widely adopted but often struggling to capture dynamic user interests and complex item relationships. Recently, deep learning techniques utilizing the attention mechanism have shown promise in modeling sequential user behavior for recommendation tasks by leveraging self-attention to identify relevant historical interactions when predicting the next item of interest. However, these models primarily rely on raw interaction sequences, failing to exploit auxiliary information, such as user ratings and item categories that could further enhance recommendation accuracy. In this study, we propose a self-attentive sequential recommendation method that enriches input representations by incorporating user ratings to capture explicit preferences and item categories to capture semantic relationships between items, extending the self-attention architecture to attend to the combination of interaction sequences, ratings, and categories. Through extensive experiments on public datasets, we demonstrate that our proposed method exhibits relative strengths in capturing certain aspects of user-item relationships, leading to competitive performance compared to original models across different recommendation quality metrics.