Recommendation Ranking Method Combining Graph Convolutional Network and Factorization Machine

With the rapid development of graph neural network technology and the wide application of personalized recommender systems in the industry, how to better apply graph representation learning to recommender systems to continuously improve the recommendation effect and improve the user experience has become a hot field of the industry research. Based on hot issues such as sparse feature data, cold start, and multi-feature combination in massive data, this paper proposes a personalized recommendation ranking method that combines graph convolutional network and factorization machine. The method represents the network relationship between users and items based on the graph structure and generates the graph embeddings on the hyperparameter space through graph representation learning, uses the factorization machine to combine the learning of four categories of features of user attributes, item attributes, interaction, and context, and generates the vector representation separately, and finally predicts personalized recommendation score based on the ranking model. The comparison experiments of multiple groups of ranking methods show that the ranking learning method proposed in this paper is better than the baseline in six indicators of AUC, Logloss, UV_CTR, CVR, ATV and the exposure ratio of new items on the online e-commerce data set, which proves the effectiveness of the proposed method.