Graph Neural Networks for Heterogeneous Trust based Social Recommendation

Abstract

In the current research, Graph Neural Networks (GNNs) play a decisive role in learning from network data structure. In a social recommender system, GNNs have a significant perspective to integrate the structure of a customer-customer social network and the customer-product bipartite network. Most of the existing trust-based social recommendation systems overlook heterogeneous trust relations among customers and heterogeneous interactions between customers and products. However, this is very challenging to capture these heterogeneous information. To address this challenge, we propose an approach to evaluate the authenticity of reviews written by customers on products. Varying authenticity introduces the heterogeneity in trust relations among customers and interactions between customers and products. This authenticity defines a customer's characteristic as a reviewer, whether the customer is reliable or biased. To the best of our knowledge, this is the first work which includes authenticity of reviews and customers to evaluate trust relationships and interactions. We develop a novel Graph Neural Network architecture for Trust-based Social Recommendation (GNNTSR) that systematically models two networks, i.e., customer-customer social network and customer-product bipartite network, and integrates heterogeneous trust and interaction. Extensive experiments are performed on real datasets, and empirical results show our model improves over the current best baseline by 2.16 - 5.74%.