Counterfactual Graph Learning for Link Prediction

Tong Zhao, Gang Liu, Daheng Wang, Wenhao Yu, Meng Jiang

Keywords: Link Prediction, Graph Representation Learning, Graph Neural Networks.

Abstract: Learning to predict missing links is important for many graph-based applications. Existing methods were designed to learn the association between two sets of variables: (1) the observed graph structure (e.g., clustering effect) and (2) the existence of link between a pair of nodes. However, the causal relationship between these variables was ignored. We visit the possibility of learning it by asking a counterfactual question: “would the link exist or not if the observed graph structure became different?” To answer this question, we leverage causal models considering the information of the node pair (i.e., learned graph representations) as context, global graph structural properties as treatment, and link existence as outcome. In this work, we propose a novel link prediction method that enhances graph learning by counterfactual inference. It creates counterfactual links from the observed ones, and learns representations from both the observed and counterfactual links. Experiments on benchmark datasets show that this novel graph learning method achieves state-of-the-art performance on link prediction.

One-sentence Summary: Design a novel method that leverages counterfactual inference to improve link prediction on graphs.

Community Implementations: [ 8 code implementations](https://www.catalyzex.com/paper/arxiv:2106.02172/code)