Graph neural network-based light-weight recommendation system

Abstract

The recommendation system aims to obtain valuable information for users and alleviate information overload. Graph Neural Network (GNN) is one the mainstream method of recommendation systems for its powerful capabilities of graph data representation and deep feature extraction. But there are still problems with the efficiency and accuracy of GNNs. Therefore, a High-Efficiency Graph Neural Network (HEGNN) is proposed in this paper to build a lightweight graph recommendation system. HEGNN strengthens the local and global preferences of users with attention blocks. It abandons the feature transformation and nonlinear activation layer of vanilla GNNs. Only the basic components are reserved to improve efficiency. Comprehensive comparative experiments with nine baseline algorithms are carried out on three benchmark datasets which include Amazon-Book Dataset, Yelp2018 Dataset, and Gowalla Dataset. Compared with existing recommendation methods such as NGCF and LightGCN, HEGCN not only achieves the highest score on two evaluation metrics of Normalized Discounted Cumulative Gain and Recall but also requires the least training time.