Visual content-enhanced sequential recommendation with feature-level attention

Abstract

A sequential recommender system (SRS) takes historical user-item interactions as a sequence to predict the next item that may be of interest to a user. Most of existing SRSs make predictions by only considering item IDs, while ignoring other important factors that may also significantly influence users’ choices, such as the visual content of items. This may not always be true in reality since an item’s appearance usually has great visual impact on users and thus plays an important role in users’ choices. In practice, each item’s appearance has often been specifically designed to represent a particular style or look. In addition, users’ preferences with regard to visual aspects, such as styling, color, etc., may be relatively consistent compared with their dynamic preferences on items for a certain period under sequential transaction scenarios. This motivates us to take the items’ visual content into account to build more reliable SRSs. Accordingly, we propose a novel visual content-enhanced sequential recommender system (VCSRS) for improving the performance of sequential recommendations. Particularly, in VCSRS, a feature-level attention module (FAM) is designed to learn the attentive visual representations of an item’s appearance. Moreover, a vision-concentrated recurrent network (VCRN) is devised to model the sequential dependencies between items while incorporating items’ visual representations. Extensive experiments on real-world datasets demonstrated the effectiveness of visual information and showed the superiority of our approach over other representative and state-of-the-art methods.