A Fusion Model of Neural Attention Mechanism and Matrix Factorization for Sequential Recommendation Systems

Abstract

Sequential recommendation systems are crucial for modern platforms that aim to provide personalized and real-time suggestions for users. By concentrating on the user’s current actions within a session or user’s history, these systems can rapidly adjust to preferences and predict the next items that the user may like. This paper introduces a method to enhance prediction accuracy by combining the Neural Attentive Recurrent Model (NARM) with Matrix Factorization (MF). The NARM utilizes its capacity to memorize and apply attention weights in sequential data, while the MF uncovers latent relationships between factors in the data. Experiments conducted on sequential datasets demonstrate that the proposed fusion model can improve the prediction results compared to the original NARM model in both Recall@20 and MRR@20 measures. This suggests that integrating latent factors of item embeddings can improve the model’s prediction accuracy.