Friend Recommendations with Self-Rescaling Graph Neural Networks

Friend recommendation service plays an important role in shaping and facilitating the growth of online social networks. Graph embedding models, which can learn low-dimensional embeddings for nodes in the social graph to effectively represent the proximity between nodes, have been widely adopted for friend recommendations. Recently, \textit{Graph Neural Networks} (GNNs) have demonstrated superiority over shallow graph embedding methods, thanks to their ability to explicitly encode neighborhood context.  This is also verified in our Xbox friend recommendation scenario, where some simplified GNNs, such as LightGCN and PPRGo, achieve the best performance. However, we observe that many GNN variants, including LightGCN and PPRGo, use a static and pre-defined normalizer in neighborhood aggregation, which is decoupled with the representation learning process and can cause the scale distortion issue. As a consequence, the true power of GNNs has not yet been fully demonstrated in friend recommendations.

In this paper, we propose a simple but effective \textit{self-rescaling network} (SSNet) to alleviate the scale distortion issue. At the core of SSNet is a generalized self-rescaling mechanism, which bridges the neighborhood aggregator’s normalization with the node embedding learning process in an end-to-end framework. Meanwhile, we provide some theoretical analysis to help us understand the benefit of SSNet. We conduct extensive offline experiments on three large-scale real-world datasets. Results demonstrate that our proposed method can significantly improve the accuracy of various GNNs. When deployed online for one month’s A/B test, our method achieves 24\% uplift in adding suggested friends actions. At last, we share some interesting findings and hope the experience can motivate future applications and research in social link predictions.