Deep Structured State Learning for Next-Period Recommendation.

Abstract

User activities in real systems are usually time-sensitive. But, most of the existing sequential models in recommender systems neglect the time-related signals. In this article, we find that users' temporal behaviors tend to be driven by their regularly changing states, which provides a new perspective on learning users' dynamic preference. However, since the individual state is usually latent, the event space is high dimensional, and meanwhile, temporal dependency of states is personalized and complex; it is challenging to represent, model, and learn the time-evolving patterns of user's state. Focusing on these challenges, we propose a deep structured state learning (DSSL) framework, which is able to learn the representation of temporal states and the complex state dependency for time-sensitive recommendation. Extensive experiments demonstrate that the DSSL achieves competitive results on four real-world recommendation datasets. Furthermore, experiments also show some interesting rules for designing the state dependency network.