Interest-based Item Representation Framework for Recommendation with Multi-Interests Capsule Network

Item representation plays an important role for recommendation, such as e-commerce, news, video, etc. It has been used by retrieval and ranking model to capture user-item relationship based on user behaviors. For recommendation systems, user interaction behaviors imply single or multi interests of the user, not only items themselves in the sequences. Existing representation learning methods mainly focus on optimizing item-based mechanism between user interaction sequences and candidate item(especially attention mechanism, sequential modeling). However, item representations learned by these methods lack modeling mechanism to reflect user interests. That is, the methods may be less effective and indirect to capture user interests. We propose a framework to learn interest-based item representations directly by introducing user Multi Interests Capsule Network(MICN). To make the framework model-agnostic, user Multi Interests Capsule Network is designed as an auxiliary task to jointly learn item-based item representations and interest-based item representations. Hence, the generic framework can be easily used to improve existing recommendation models without model redesign. The proposed approach is evaluated on multiple types of benchmarks. Furthermore, we investigate several situations on various deep neural networks, different length of behavior sequences and joint learning ratio of interest-based item representations. Experiment shows a great enhancement on performance of various recommendation models and has also validated our approach. We expect the framework could be widely used for recommendation systems.