E-Commerce Purchase Prediction Based on Graph Neural Networks

Abstract

This With the booming development of the Internet, many online trading platforms have emerged, and e-commerce shopping is gradually becoming an important way of consumption for people. While e-commerce brings convenience, it also brings the problem of information overload. Ineffectively selecting products at sea wastes time and reduces the efficiency of product distribution. Therefore, product recommendation has become an important way to alleviate the information overload of e-commerce shopping. In this paper, we formulate the problem as a multi-label prediction task and propose a new graph neural network-based framework, the item-relational graph neural network (IRGNN), for the simultaneous discovery of multiple complex relationships. The feature vectors of nodes are encoded and learned based on known network structures, and the learned node features are decoded to explore the statistical patterns of node features and link formation between nodes, and to analyse which features are prone to form links between nodes that possess them. Extensive experiments have been conducted on real-world product data to verify the effectiveness of IRGNN, particularly on large and sparse product graphs.