Integration of Multiple Time Embedding and GLU for Sequential Recommendation

Abstract

Self-attention has made great progress in the field of sequential recommendation. The traditional attention-based sequential recommendation model usually adopts position embedding or simple time embedding, which is difficult to use the user’s time information to obtain the changes between the long-term and short-term interests of users, which makes the sources of user interest information single and insufficient. In order to solve the above problems, the self-attention sequential recommendation algorithm is proposed, which combines with multiple timing characteristic embedding and gated linear feed-forward network. The feed-forward neural network with gated linear units (GLU) is used to better optimize the role of transformer model in sequential recommendation, and multiple time tag embedding methods are used to fully obtain the changing trend of user interest over time, so as to improve the accuracy of recommendation. Experiments on multiple datasets indicate that we solve the problems better and improve the predictive performance of the model compared with the traditional model.