Temporal-order association-based dynamic graph evolution for recommendation

Modeling the interactions between users and items to accurately predict a user preference on items is very crucial for improving the performance of recommendation. Although existing graph-based methods have achieved great progress in predicting a user preference for recommendation, they usually need additional side information which is difficultly obtained, and ignore the temporal-order associations between items (users) when constructing graphs. In this paper, we propose a temporal-order association-based dynamic graph evolution model for recommendation, which can not only capture temporal-order item associations and user relationships by recurrently constructing a temporal-order item association graph and a user similarity graph but also update and promote the embeddings and graphs by performing a novel dynamic evolution mechanism based on two graph convolutional networks (GCNs). Specifically, the proposed model consists of two main components: recurrent graph construction component and message propagation and aggregation component. The former recurrently constructs the temporal-order item association graph and the user similarity graph only from the history interactions and embeddings to capture the item-item and user-user relationships. The latter performs a novel dynamic evolution mechanism based on two GCNs on these two auxiliary graphs and interaction graph to refine user and item representations, which further helps the process of constructing two auxiliary graphs. Finally, the final embeddings of users and items are used to predict a user preference on all items the user has not interacted with. The experimental results illustrate that our method outperforms the state-of-the-art methods on five real-world datasets.

Automated summary

In this paper, a dynamic graph evolution model is proposed to accurately predict user preference. The model captures temporal-order item associations and user relationships, and updates embeddings and graphs through a dynamic evolution mechanism based on two GCNs. Experimental results show that the proposed method outperforms state-of-the-art methods on real-world datasets.