An Explainable Recommender System Based on Heterogeneous Information Networks and User's Reviews

Abstract

In the realm of explainable recommendations, the primary goal is not only to provide users with item recommendations but also to offer insights into why specific items are being suggested. Traditional recommendation methods rely on user-item rating data to model user preferences, but their effectiveness is limited due to data sparsity. To overcome these challenges, our approach leverages the power of heterogeneous information networks (HIN) and textual reviews to enhance recommendation performance and transparency, aiming to provide users with clear explanations for each recommendation. In this paper, we introduce AEHIN, an innovative attention-based recommender system. AEHIN enhances the recommendation process by thoughtfully selecting relevant meta-paths for user and item feature extraction. It employs matrix factorization in conjunction with an attention mechanism to evaluate the importance of each meta-path. Furthermore, AEHIN incorporates review texts, leveraging a temporal convolutional network (TCN) language model with dual attention mechanisms to improve the extraction of user and item features from the review content. The integration of HIN and review texts is facilitated through the attention mechanism, enhancing recommendation transparency and model interpretability. To predict user ratings for items, a standard multi-layer perceptron is used, which captures the nonlinear interactions between users and items.

Keywords: Recommender systems, heterogeneous information networks, Attention Mechanism, temporal convolutional network, review texts, explainability