Discovering Proper Neighbors to Improve Session-Based Recommendation

Abstract

Session-based recommendation shows increasing importance in E-commerce, news and multimedia applications. Its main challenge is to predict next item just using a short anonymous behavior sequence. Some works introduce other close similar sessions as complementary to help recommendation. But users’ online behaviors are diverse and very similar sessions are always rare, so the information provided by such similar sessions is limited. In fact, if we observe the data at the high level of coarse granularity, we will find that they may present certain regularity of content and patterns. The selection of close neighborhood sessions at tag level can solve the problem of data sparsity and improve the quality of recommendation. Therefore, we propose a novel model CoKnow that is a collaborative knowledge-aware session-based recommendation model. In this model, we establish a tag-based neighbor selection mechanism. Specifically, CoKnow contains two modules: Current session modeling with item tag(Cu-tag) and Neighbor session modeling with item tag (Ne-tag). In Cu-tag, we construct an item graph and a tag graph based on current session, and use graph neural networks to learn the representations of items and tags. In Ne-tag, a memory matrix is used to store the representations of neighborhood sessions with tag information, and then we integrate these representations according to their similarity with current session to get the output. Finally, the outputs of these two modules are combined to obtain the final representation of session for recommendation. Extensive experiments on real-world datasets show that our proposed model outperforms other state-of-the-art methods consistently.