Cross-domain recommendation via adaptive bi-directional transfer graph neural networks

Yi Zhao, Jingxin Ju, Jibing Gong, Jinye Zhao, Mengpan Chen, Le Chen, Xinchao Feng, Jiquan Peng

Abstract: Data sparsity and the cold start problem significantly impede the advancement of recommendation systems. Cross-domain recommendation (CDR) seeks to alleviate these issues by utilizing knowledge from other domains. One important research direction in CDR is how to effectively transfer knowledge across different domains. Existing methods assume that knowledge between different domains should be transferred among the same users across domains. However, a user may share similar interests with multiple distinct users in other domains, which means that the domain knowledge from one domain should not only be transferred to the same user in another domain but should also be extended to other users with similar interests in that domain. Additionally, a user may experience a phenomenon of interest drift across different domains. The prior assumption often hampers the effective dissemination of domain knowledge, thus limiting the performance of CDR. Therefore we propose AbtCDR, an innovative method that enables adaptive knowledge transfer among different users in different domains. Firstly, we harness the power of graph neural networks to distill user and item embeddings across different domains. Secondly, we define a novel cross-domain user interest transport scheme, meticulously designed to bridge the gap between distinct domains. To further navigate the complex web of user interests, we have proposed two knowledge transfer methods: a coarse-grained method for broad-strokes insight into user interests, and a fine-grained method that delves into the intricate details of user interests. Extensive experiments demonstrate our method outperforms state-of-the-art baselines on real-world datasets. Our code and dataset are available at https://github.com/jujingxin/AbtCDR.

External IDs:dblp:journals/kais/ZhaoJGZCCFP25