Modeling User Interest Changes with Dynamic Differential Graphs for Item Recommendation

Abstract

User interests are significant components in recommendation systems. Modeling user interests based on users' historical behaviors is a challenging problem, and many recommendation models have been proposed for user interests modeling, such as long-term and short-term interests modeling. In the real world, users' interests always change over time, however, existing models rarely consider users' interest changes. The purpose of this research is to apply graph neural networks to capture users' interest changes. This research first conducts data analysis on two public datasets, and results show that there are considerable amounts of users with a trend of interest changes. Based on this analysis, we construct user-category dynamic differential graphs, and we design a novel neural network based on dynamic differential graphs to learn users' interest changes representations from dynamic differential graphs. The learned representations are integrated with long-term and short-term interest representations to get users' final representations and make recommendations by getting scores with items. Different types of experiments are conducted to evaluate the performance of our proposed model, and experiment results show that the proposed model outperforms other baseline models.