Aspect-Based Capsule Network With Mutual Attention for Recommendations

Abstract

Review text is a valuable source of information for recommendation systems and often contains rich semantics with user preferences and item attributes. Recently, mainstream recommendation approaches have been using deep learning techniques to utilize review text. These efforts employ an affinity matrix to model the correlations between users and items and in turn aggregate the contextual features of users and items to form latent representations with some interpretability. However, constructing this simple interaction does not capture the complex correlations between users and items. Furthermore, a single representation of latent features is not sufficient to express user preferences and item attributes in reviews with multiple granularities. To this end, we propose an aspect-based capsule network and mutual attention for recommendation systems, named Acapnet. We design a novel contextual mutual attention mechanism for modeling the fine-grained correlations in the contextual features of users and items. We employ reverse dynamic routing to aggregate user and item context features into aspect features (capsules) for rating predictions; the vector length of a capsule can be used as a basis for estimating aspect importance. Extensive experiments are conducted on five real-world datasets with different features. The results show that the proposed Acapnet is superior to recent state-of-the-art models in terms of performance. Further analysis reveals that our model effectively mitigates the homogeneity among various aspect capsules with strong interpretability. <p xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Impact Statement—On e-commerce platforms such as Amazon and Yelp, recommendation systems play a key role as a filtering technology that helps consumers find the best product or service among many choices. However, recent studies have shown that although it can provide relatively accurate services, it often fails to explain why that product is recommended to the user. The model we propose in this article, Acapnet, effectively overcomes the limitation. Compared to state-of-the-art recommendation methods, Acapnet not only outperforms in terms of performance, but also offers some superiority on interpretability.