GraphRR: A multiplex Graph based Reciprocal friend Recommender system with applications on online gaming service

Abstract

Reciprocal Recommender Systems (RRSs) are recommender systems specifically designed for people-to-people recommendation tasks, e.g., online gaming, dating, and recruitment services. They are fundamentally different from the conventional user–item recommendations. In RRSs, user interactions are usually directional, i.e., they are initiated by one side and not necessarily reciprocated by the other side. In the meanwhile, abundant multiplex user interactions, e.g., Friend Request and Send Message, are collated by the online services and can be represented into a large-scale multiplex user interaction graph. Despite the substantial progress of Graph Neural Networks (GNNs) on capturing users’ multiplex interactions, naive GNNs are insufficient to capture the additional information implied from the directions of interactions, as they are usually not designed to preserve the asymmetric proximities between users.In the paper, we present a novel Graph neural network for Reciprocal Recommendation (GraphRR) to utilize the multiplex user interactions. Specifically, three ego graphs are augmented based on the directions of interactions for each user to capture his preference, attraction and similarity in a finer granularity. Then the multiplexity-aware GNN modules are further applied to measure the contributions of different interaction types. Extensive experiments are conducted in the datasets of the real-world large-scale online games from NetEase Games, a leading game provider for worldwide users. The experimental results demonstrate the superiority of GraphRR over baseline methods and provide empirical evidence for the benefits of the proposed ego graph augmentation. The source code is also available online for reproductivity11https://github.com/changym3/GraphRR..