DeepInteract: Multi-view features interactive learning for sequential recommendation

Abstract

Deep learning models have been successfully applied in sequential recommendations. However, previous studies ignored the interaction between static and dynamic features of both items and users, thus fail to exactly capture users’ current preferences. To overcome this limitation, we first conducted feature representations from static, dynamic and interactive views and constructed corresponding feature mining modules. Then, based on the multi-view feature mining modules, we proposed a deep learning-based model, namely DeepInteract, to learn the interaction of multi-view features of both item profiles and user behaviors for sequential recommendation. Experimental results on three real-world datasets demonstrated that DeepInteract outperforms state-of-the-art methods not only on recommendation performance but also on stability and robustness. Furthermore, we used ablation experiments to investigate the importance of three feature mining modules on various measures of recommendation performance. It was demonstrated that interactive features play the most important role for sequential recommendation.