Dynamic item feature modeling for rating prediction in recommender systems

Conventional recommendation methods focus on optimizing user and item representations using various modelling methods. While persistent features of items are well studied, time varying hidden features of items are largely neglected. We argue that it is desirable to model both static and dynamic representations of items in one framework. Moreover, dynamic features often exhibit periodic variation characteristics. Identifying dynamic features of items can help merchants recognize evolving trends of their product and provide better services to customers for more trading benefit. Based on our observations, a period-aware correlational-temporal user/item feature modeling method in the form of a double-chained BiGRU model with attention mechanism is proposed. Furthermore, heterogeneous graph-based meta-paths are incorporated to model static features of items. To the best of our knowledge, this is the first effort to model both static and dynamic representations of items in one setting. A heterogeneous correlational temporal framework (HCoTemp) fusing static and dynamic item representations along with dynamic user representation for sequential recommendation is proposed. Empirical studies of 4 Amazon review benchmark datasets demonstrate that our model outperforms state-of-the-art methods in both MSE and HR. We also conducted extensive ablation experiments, which reveal that each component of HCoTemp contributes to performance improvements. Randomly selected cases from the Amazon Game dataset also confirm our findings.