A cross-domain recommendation model by unified modelling high-order information and rating information

Cross-domain recommendation models are proposed to enrich the knowledge in the target domain by taking advantage of the data in the auxiliary domain to mitigate sparsity and cold-start user problems. However, most of the existing cross-domain recommendation models are dependent on rating information of items, ignoring high-order information contained in the graph data structure. In this study, we develop a novel cross-domain recommendation model by unified modelling high-order information and rating information to tackle the research gaps. Different from previous research work, we apply heterogeneous graph neural network to extract high-order information among users, items and features; obtain high-order information embeddings of users and items; and then use neural network to extract rating information and obtain user rating information embeddings by a non-linear mapping function MLP (Multilayer Perceptron). Moreover, high-order information embeddings and rating information embeddings are fused in a unified way to complete the final rating prediction, and the gradient descent method is adopted to learn the parameters of the model based on the loss function. Experiments conducted on two real-world data sets including 3,032,642 ratings from two experimental scenarios demonstrate that our model can effectively alleviate the problems of sparsity and cold-start users simultaneously, and significantly outperforms the baseline models using a variety of recommendation accuracy metrics.