Large-scale online multi-view graph neural network and applications

Abstract Recently popularized Graph Neural Network (GNN) has been attaching great attention along with its successful industry applications. This paper focuses on two challenges traditional GNN frameworks face: (i) most of them are transductive and mainly concentrate on homogeneous networks considering single typed nodes and edges; (ii) they are difficult to handle the real-time changing network structures as well as scale to big graph data. To address these issues, a novel attention-based Heterogeneous Multi-view Graph Neural Network (aHMGNN) solution is introduced. aHMGNN models a more intricate heterogeneous multi-view network, where various node and edge types co-exist and each of these objects also contain specific attributes. It is end-to-end, and two stages are designed for node embeddings learning and multi-typed node and edge representations fusion, respectively. Experimental studies on large-scale spam detection and link prediction tasks clearly verify the efficiency and effectiveness of our proposed aHMGNN. Furthermore, we have implemented our approach in one of the largest e-commerce platforms which further verifies that aHMGNN is arguably promising and scalable in real-world applications.