Attention-Based User Temporal Model for Recommendation

Abstract

For the traditional recommendation system, the user model is represented by the user’s historical behaviors, but the user’s interest is not static, and the user’s interest drift affects the final recommendation effect of the recommendation system. However, most of the work to solve the drift of interest is to design long-term and short-term models for users, Using RNNbased models with different time spans to learn an overall embedding of user history sequences. It is not possible to dynamically monitor user interest drift. In this work, we propose a user model named Attention-Based User Temporal Model (AUTM). The key design of our work is to use the attention model to dynamically assign importance weights to user embedding of RNN network outputs at different times, and then combine the static user embedding. Compared to a set of baselines in real-word dataset, our model shows better performance in prediction precision and Area Under Curve (AUC) score.