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 reproductivity¹.

Introduction

Recommender Systems (RSs) have been widely developed and employed in a variety of applications including E-commerce shopping systems [1], [2], social media [3], and online video service [4], etc. However, these conventional user–item RSs are incapable of accommodating people-to-people recommendation scenarios, as they only consider users’ unidirectional preference toward inanimate items. While the recommended items cannot “reject” users in the classical RSs, unilateral behaviors occur frequently in the people-to-people recommendation.

Reciprocal Recommender Systems (RRSs) aim to connect users with mutual preference, i.e., reciprocity, in the people-to-people recommendation scenarios. For instance, there are some RRSs in online games designed for recommending mutually preferable teammates/friends to increase gamer’s engagement. Fig. 1 shows an example of RRS on the online game Knives Out, which is a popular mobile game in China. When gamers require the recommendation result, the recommender system aims to provide a user list that is appealing to users from both sides. Apart from the online gaming service, there also exist many other important domains that involve reciprocal recommendations, including online social platform [5], online dating [6] and recruitments [7], etc. In spite of the urgent requirements of these applications, RRSs have received little attention compared to the conventional recommender system.

The objective of RRSs is to create a mutually preferable match between two users, which renders the task of reciprocal recommendation more complex and challenging. The RRS research in the literature can be generally divided into two categories: Content-Based (CB) and Collaborative Filtering (CF). The CB approaches [6] are primarily based on the description of items and user profiles, thus heavily relying on the attributes by feature engineering. The CF-based approaches [8], [9] propose to capture the collaborative signals between similar users based on their historical interactions. The limitation of CF-based methods lies in their non-scalability due to the extravagant calculation of the similarities over all pairs of users. A recent work [10] takes advantage of the latent factor model to generate user embeddings from the dual interaction matrix, showing significant improvement on both performance and computational efficiency compared to other CF-based approaches. Nevertheless, it is still limited in dating scenarios, which inherently have two clear roles (male and female) to be matched. It cannot generalize well to some complex reciprocal scenarios, e.g., the online gaming service.

As user interactions can be intrinsically described in graph-structured data, where nodes represent users/items and edges represent users’ interactive behaviors, there have been many prominent studies [3], [11], [12] incorporating graph-based techniques into recommender systems. These methods leverage Graph Neural Networks (GNNs) [13], an emerging paradigm for graph representation learning, to effectively generate user representations. One of the significant advantages is their expressive capacities to jointly learn node attributes and graph structures. Another advantage lies in their strong extensibility to tackle the multiplexity of graphs, where multiple types of edges may exist between two nodes. Recently, some works [14], [15] have also demonstrated the effectiveness of multiplex GNNs in the user interaction graph, facilitating the recommender systems to explore more fine-grained user portraits.

Despite their promising success in conventional recommender systems, GNNs remain relatively unexplored in the large-scale reciprocal recommender system. It is beneficial to combine GNNs’ capacity into RRS to determine users’ reciprocity. However, designing GNNs for reciprocal recommender systems poses unique challenges. First, the requirement of reciprocity cannot be directly fulfilled by applying GNNs on the directional interaction graph as in the conventional user–item recommendation scenarios. We will show a detailed analysis in the paper that user interaction graphs declare the properties of heterophily, asymmetry, and non-transitivity. These properties render traditional GNNs insufficient to capture the information implied from the directions of interactions. Second, users continually generate multiplex interactions, which reveal different aspects of user portraits in the online service. The measurement of contributions of multiplex interaction needs to be considered by the RRS.

Concerning the challenges mentioned above, in this work, we propose a carefully designed multiplex Graph based Reciprocal friend Recommender system (GraphRR) with applications on two online games to enhance gamers’ engagement. Here we propose five research questions to guide readers to introduce our works and contributions:

• Why common recommendation methods are insufficient to capture users’ portraits?

• Why multiplexity is important in the scenarios of recommendation?

• Why does the direction of each interaction need to be considered in the reciprocal recommendation?

• Why GNNs cannot be directly applied in the directional user interaction graph for the reciprocal recommendation?

• How to utilize GNNs in the task of reciprocal recommendation?

To summarize, our key contributions are three-fold:

• In view of the unidirectionality of the user interaction graph, we analyze its property of heterophily, asymmetry, and non-transitivity, which is handled by the proposed mechanism of the ego graph augmentation.

• We provide a GNN-based solution for RRSs, namely GraphRR, to capture users’ preference, attraction and similarity separately. Besides, the attentive aggregation is integrated into GraphRR to further utilize the multiplex users’ interactions in the online service.

• We conduct extensive experiments on the datasets from real-world online games and demonstrate that the proposed GraphRR outperforms a series of baseline methods.

The remainder of the paper is structured as follows: Section 2 discusses the related work. Section 3 formulates the task of the reciprocal recommendation and describes the notations used in the paper. In Section 4, we provide the analyses that motivate the designs of the proposed GraphRR. In Section 5, we introduce the details of the proposed GraphRR, including the ego-graph augmentation and the multiplex graph neural networks. Experimental results are shown in Section 6 to verify the effectiveness of the proposed method. Finally, the conclusion and future work are discussed in Section 7.