Knowledge Graph-Based Personalized E-commerce Product Recommendations: Aligning with Consumer Interests

Large-scale e-commerce platforms often encompass various recommendation scenarios to cater to the diverse needs of different consumers. In recent years, the application of knowledge graph-based multi-domain recommendation (MDR) has drawn increasing attention of researchers. The state-of-the-art MDR methods have typically shared structures, shared parameters, and domain-specific components with specific parameters. However, due to the variations in data distributions across different domains, shared parameters often suffer from domain conflicts during the training process, while domain-specific parameters are prone to overfitting in sparse data domains. To address these problems, we have proposed the GDKG2D (GDKG+DN+DR) model for multi-domain recommendation of products. An unsupervised learning approach is firstly employed to construct a generalized domain knowledge graph (GDKG), in which a Local Weighted Smoothing (LWS) scheme according to consumer interests has been designed to encode the nodes. Node embeddings is then used in LWS to train encoders for consumers in the consumer-product bipartite graph. Subsequently, a Domain Negotiation (DN) strategy is employed to alleviate domain conflicts. Domain Regularization (DR) is applied finally in training in other domains to enhance the generalization performance of specific parameters. After evaluating on the product recommendation datasets from Amazon and Taobao, the GDKG2D model demonstrates significant advantages over conventional MDR methods. It exhibits superior performance, particularly in terms of AUC scores.