CaSe4SR: Using category sequence graph to augment session-based recommendation

Abstract

Session-based recommendation aims to predict next item based on users’ anonymous behavior sequence within a short time. Recent studies focus on modeling sequential dependencies or complex relations among items in a session via recurrent/convolutional/graph neural networks. However, the following problems still remain: for short sessions, limited interactions cannot manifest user’s intent clearly; for long sessions, user’s interest may drift but be blurred by complex transitions. Motivated by the observation that different items are often belong to only a few categories or that closely related, in this article, we tackle these challenges by leveraging item category information, which is a concise form of knowledge and readily available in many platforms. We propose a novel method CaSe4SR that utilizes category sequence graph to augment session-based recommendation. In CaSe4SR, we build an item graph and a category graph, from user behavior sequence and item category sequence. The latter summarizes the former at concept level, which reduces item-level user behavior noises and makes user’s interest clearer. Afterwards, graph neural networks are applied on item graph and category graph respectively to learn representations of items and categories. Then two alternative fusion strategies and attention mechanism are designed to integrate them, yielding global embedding of the session, which is further combined with representation of last item to get ultimate session representation. Extensive experiments on real-world datasets show that CaSe4SR outperforms other state-of-the-art methods consistently. Detailed analysis reveals that category sequence graph is beneficial for next-item recommendation in sessions with different lengths.