Aspect-level recommendation fused with review and rating representations

Abstract

Review contains user opinions about different aspects of an item, which is essential data for aspect-level recommendation. Most existing aspect-level recommendation algorithms are concerned with the degree to which user and item aspects match. However, even if an item is extremely popular due to its high quality, it may only partially match the aspects of a user. A tolerant user may like the item, whereas a strict user may dislike it. This implies that these works disregard the personalized behavior patterns of the user. In this paper, we propose a new Aspect-level Recommendation model fused with Review and Rating, namely ARRR, to address the recommendation bias. First, we introduce rating to explore user behavior patterns and item quality. Then, we present a personalized attention mechanism that generates a set of aspect-level user or item representations from reviews. Finally, we obtain comprehensive user or item representations by combining rating- and review-based representations. In the experiments, the proposed model is compared with seven state-of-the-art recommendation algorithms on seven datasets. The results show that our model outperforms on seven metrics. The source code of ARRR is available at https://github.com/alinn00/ARRR.