Graph factorization machine based recommendation algorithm with graph construction and attention mechanism

Abstract

Graphs, as a type of data structure that exists in numerous scenarios, are applied in various fields of the internet, such as recommendation algorithms, community detection, path search, and so on. In recent years, with the rapid development of the internet, the amount of graph data has greatly increased, and analyzing and utilizing these graph data to establish graph network models and apply them to various real-life production and living scenarios is of significant importance, such as using user-item data to establish a recommendation algorithm model for recommendations. Currently, many graph neural network models have achieved good results, but there is still room for improvement. Better algorithms can more accurately understand the user's needs and bring a better recommendation experience to the user. The paper designs a new way of constructing graphs and improves the graph network message-passing algorithm. A new graph neural network algorithm, GAFM (Graph Attention Factorization Machines), is proposed. Compared with traditional algorithms, this algorithm will construct graph structures of users and items based on original data, and can also construct item graph data by linking to a knowledge base to introduce new items. When this algorithm performs message passing between nodes, it will consider the first-order neighbor set of the central node and the second-order cross-item of the first-order neighbor set, and use an attention mechanism to regulate the weight in the message passing process to capture high-dimensional neighbor information and improve accuracy. The experimental results on real-world datasets show that the algorithm has better performance compared to existing graph recommendation algorithms.