Enhanced Multi-Relationships Integration Graph Convolutional Network for Inferring Substitutable and Complementary Items

Understanding the relationships between items can improve accuracy and interpretability of recommender systems. Among these relationships, substitute complement attract most attention in e-commerce platforms. The substitutable are interchangeable might be compared with each other before purchasing, while complementary used conjunction usually bought together query item. In this paper, we focus on two issues inferring items: 1) how to model their mutual influence performance downstream tasks, 2) further discriminate them by considering strength relationship for different item pairs. We propose a novel multi-task learning framework named Enhanced Multi-Relationships Integration Graph Convolutional Network (EMRIGCN). regard inference task as link prediction heterogeneous graph types edges nodes (items). To complement, EMRIGCN adopts two-level integration module, i.e., feature structure integration, based experts sharing mechanism during message passing. obtain pairs, build an auxiliary loss function increase or decrease distances embeddings weak strong relation latent space. Extensive experiments both public industrial datasets prove that significantly outperforms state-of-the-art solutions. also conducted A/B tests real world systems Meituan Maicai, online supermarket platform China, obtained 15.3% improvement VBR 15.34% RPM.