Capturing Multi-granularity Interests with Capsule Attentive Network for Sequential Recommendation

The sequential recommender system attempts to predict the next interaction based on user’s historical behaviors, which is a challenging problem due to intricate sequential dependencies and user’s various interests underneath the interactions. Existing works regard each item that the user interacts with as an interest unit and apply advanced deep learning techniques to learn a unified interest representation. However, user’s interests vary in multiple granularities. An item mirrors preferences for a specific item, while a set of items reflect general user interests, which are barely captured by a unified representation at the same granularity level. Furthermore, since the unrelated items are treated at the same granularity level as these decisive items, the model cannot focus on the items that help make the accurate recommendation. In this paper, we propose a novel Capsule Attentive Network (CAN) for sequential recommendation, which integrates the dynamic routing algorithm to capture diverse user interests at coarse-grained levels with a transformer module to learn more informative embeddings at fined-grained levels. Experimental results on three datasets demonstrate that CAN achieves substantial improvement over state-of-the-art methods.