Graph-Based Sequential Interpolation Recommender for Cold-Start Users

Abstract

Sequential recommendation systems aim to capture the users’ dynamic preferences from users’ historical interactions to provide more accurate recommendations. However, in many scenarios, there are a large number of cold-start users with limited user-item interactions. To address this challenge, some studies utilize auxiliary information to infer users’ interests. But with the increasing awareness of personal privacy protection, it is difficult to obtain detailed descriptions of users. Therefore, we propose a model called GitRec (Graph-based sequential interpolation Recommender for cold-start users) to address this issue. Our proposed model GitRec captures users’ latent preferences and provides more proper recommendations for cold-start users without any help from external information. Specifically, (i) GitRec constructs the global graph based on all sequences to obtain the core interests of the warm-start users. (ii) Then, GitRec selects content-rich sequences that are most similar to the short sequences and constructs the local graph to find out the potential preferences of cold-start users. (iii) Finally, GitRec can find suitable candidates to enrich the information of the cold-start users by using graph neural networks, and provide better recommendations with the help of core interests and potential preferences. We conducted extensive experiments on three real-world datasets. The experimental results demonstrate that GitRec has significant improvement over state-of-the-art methods.