[Paper Review] Counterfactual Graph Learning for Link Prediction.

Tong Zhao, Gang Liu+3 more|arXiv (Cornell University)|Jun 3, 2021

Advanced Graph Neural Networks76 references4 citations

TL;DR

This paper proposes a novel link prediction method that leverages counterfactual inference to improve graph representation learning by generating synthetic counterfactual links from observed ones. By modeling node pair context as context, global graph structure as treatment, and link existence as outcome, the method learns enhanced representations from both real and counterfactual links, achieving state-of-the-art performance on benchmark datasets.

ABSTRACT

Learning to predict missing links is important for many graph-based applications. Existing methods were designed to learn the observed association between two sets of variables: (1) the observed graph structure and (2) the existence of link between a pair of nodes. However, the causal relationship between these variables was ignored and we visit the possibility of learning it by simply asking a counterfactual question: would the link exist or not if the observed graph structure became different? To answer this question by causal inference, we consider the information of the node pair 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 the counterfactual inference. It creates counterfactual links from the observed ones, and our method learns representations from both of them. Experiments on a number of benchmark datasets show that our proposed method achieves the state-of-the-art performance on link prediction.

Motivation & Objective

• To address the limitation of existing link prediction methods that ignore causal relationships between graph structure and link existence.
• To investigate whether counterfactual reasoning can improve link prediction by simulating how link existence would change under altered graph structures.
• To develop a method that jointly learns from observed and counterfactual links to produce more robust node representations.
• To demonstrate that incorporating counterfactual inference leads to superior performance on standard link prediction benchmarks.

Proposed method

• The method models link prediction as a counterfactual inference problem, treating node pair context as context, global graph structure as treatment, and link existence as outcome.
• It generates counterfactual links by perturbing the observed graph structure while preserving node-level features and structural context.
• The model learns node representations using both observed and counterfactual links through a unified graph neural network framework.
• Counterfactual links are constructed by altering structural properties (e.g., edge presence) and evaluating how link existence would change under such interventions.
• The framework integrates counterfactual learning into the message-passing mechanism of graph neural networks to improve representation quality.
• The final link prediction is performed using a decoder that leverages the enhanced representations from both real and counterfactual link signals.

Experimental results

Research questions

• RQ1 Can counterfactual reasoning improve link prediction performance by simulating how link existence would change under modified graph structures?
• RQ2 How does incorporating counterfactual links affect the quality of learned node representations compared to using only observed links?
• RQ3 To what extent does the proposed method outperform existing state-of-the-art link prediction baselines on standard benchmark datasets?
• RQ4 What is the contribution of global graph structural properties as treatment in the counterfactual inference setup for link prediction?

Key findings

• The proposed method achieves state-of-the-art performance on multiple benchmark link prediction datasets by integrating counterfactual inference into graph learning.
• The inclusion of counterfactual links leads to more robust and generalizable node representations compared to models trained only on observed links.
• Experiments confirm that counterfactual reasoning enhances model performance by capturing latent causal dependencies between graph structure and link existence.
• The method demonstrates consistent improvements across diverse graph types and link prediction settings, validating its generalization capability.