HGNNDR: Heterogeneous Graph Neural Network Recommendation Model based on Feature Aggregation

In recent years, with the rapid development of information technology such as big data, cloud computing, and the Internet of Things, the scale of data contained in the Internet has been growing explosively. The arrival of the big data era will bring more convenience to people. However, it will also bring people the problem of "information overload", that is, when people face a huge amount of information data, many redundant and useless data will be lost. Heterogeneous graphs have diverse and complex types of data structures, how to mine information effectively for heterogeneous graphs, whether the embedding of nodes contains neighborhood information and semantic information, and so on, are all challenges in the representation learning of heterogeneous graphs. In order to address the above shortcomings, this paper combines graph neural networks with heterogeneous networks to propose the deep recommendation model HGNNDR. In this paper, a multi-feature joint representation of information in heterogeneous information networks is given. Oriented to the two domains of user/item interaction and user socialization, the multi-order topology information in heterogeneous information networks is utilized to enhance the representation of node features in the case of sparse network connections. A graph neural network recommendation method oriented to the attention mechanism is proposed. Through the effective fusion of multidimensional representation vectors of users, items, ratings, and socialization, the recommendation quality of the recommendation model is enhanced in the case of sparse rating matrices. The recommended accuracy of the model in this paper is verified to be better than the baseline method on multiple public datasets.