Heterogeneous graph neural networks with neighbor-SIM attention mechanism for substitute product recommendation

Substitute product recommendation is important to improve customer experience in E-commerce. Recent developments in Graph Neural Networks (GNNs) show superior advantages in recommendation systems, as they are able to leverage both content information of each item, and connections among items that reflect customer behavior interactions. Despite a substantial amount of effort has been made to homogeneous and heterogeneous GNN, few works shown generalization and efficiency in large-scale heterogeneous graph structure and none of them are optimized for substitute product recommendation task. In this paper, first, we present HetSAGE, a novel general framework for heterogeneous graph structure. In HetSAGE, we build a two-level hierarchical information aggregation: neighbor-level aggregation and edge-level aggregation. Specifically, for each node, the neighbor-level aggregator aims to aggregate information from its neighbor nodes connected with the same type of edges, and the edge-level aggregator further aggregates different types of edges. Second, we propose a novel Neighbor-SIM attention mechanism for edge-level aggregation, which is optimized for substitute product recommendation task. Extensive experiments on both proprietary and public Amazon datasets illustrate the best performance of HetSAGE with Neighbor-SIM attention mechanism compared with state-of-the-art GNN methods on the substitute product recommendation tasks.