Multi-Level Coupling Network for Non-IID Sequential Recommendation

Sequential recommendation has been recently attracting a lot attention to suggest users with next items to interact. However, most of the traditional studies implicitly assume that users and items are independent and identically distributed (IID) and ignore the couplings within and between users and items. Although deep learning techniques allow the model to learn coupling relationships, they only capture a partial picture of the underlying non-IID problem. We argue that it is essential to distinguish the interactions between multiple aspects at different levels explicitly for users’ dynamic preference modeling. On the other hand, existing non-IID recommendation methods are not well designed for the sequential recommendation task since users’ interaction sequence causes more complicated couplings within the system. Hence, to systematically exploit the non-IID theory for coupling learning of sequential recommendation, we propose a non-IID sequential recommendation framework, which extracts users’ dynamic preferences from complex couplings between three aspects, including users, interacted items and target items at feature level, entity level and preference level. Furthermore, to capture the couplings effectively, we implement the framework on the base of capsule network. We indicate that the basic ideas of capsule seamlessly suit our purpose of modeling sophisticated couplings. Finally, we conduct extensive experiments to demonstrate the effectiveness of our approach. The results on three public datasets show that our model consistently outperforms six state-of-the-art methods over three ranking metrics.