Improving current interest with item and review sequential patterns for sequential recommendation

Sequential recommendation (SR) aims to recommend items based on user information and behavior sequences. Almost all the existing works for SR construct short-term preference and long-term preference only based on the user-item interactions or the reviews rather than considering the two types of information simultaneously. In fact, interaction items and reviews commonly reflect the user's semantic information, and play significant roles in modeling the user preference. In this paper, we propose a novel model named Parallel Item sequential pattern and Review Sequential Pattern (PIRSP) for the sequential recommendation. Specifically, first, PIRSP learns two levels of sequential patterns from item and review information, respectively: (1) item sequential pattern, which uses a gated recurrent unit with an item-attention mechanism to model history behavior sequences; (2) review sequential pattern, which takes a convolution neural network with a target-attention mechanism for modeling associated reviews of interaction items. Then, we introduce a fusion gating mechanism for selectively combining the two sequential patterns to learn the short-term preference. Second, we employ a convolution neural network with aspect information to learn the long-term preference. Finally, we utilize a linear fusion on the long-term preference and short-term preference for modeling user preference and making final recommendation. The experimental results indicate that our model outperforms other state-of-the-art methods on the Amazon dataset. Our analysis of PIRSP's recommendation process shows the positive effect of the two types of information and fusion gating mechanism on the performance of sequential recommendation.