Joint Modeling Dynamic Preferences of Users and Items Using Reviews for Sequential Recommendation

The emerging of sequential recommender (SR) has attracted increasing attention in recent years, which focuses on understanding and modeling the temporal dynamic of user behaviors hidden in the sequence of user-item interactions. However, with the tremendous increase of users and items, SR still faces several challenges: (1) the hardness of modeling user interests from spare explicit feedback; (2) the time and semantic irregularities hidden in the user’s successive actions. In this study, we present a neural network-based sequential recommender model to learn the temporal-aware user preferences and item popularity jointly from reviews. The proposed model consists of the semantic extracting layer and the dynamic feature learning layer, besides the embedding layer and the output layer. To alleviate the data sparse issue, the semantic extracting layer focuses on exploiting the enriched semantic information hidden in reviews. To address the time and semantic irregularities hidden in user behaviors, the dynamic feature learning layer leverages convolutional fitters with varying size, integrating with a time-ware controller to capture the temporal dynamic of user and item features from multiple temporal dimensions. The experimental results demonstrate that our proposed model outperforms several state-of-art methods consistently.