[Paper Review] Counterfactual Learning on Graphs: A Survey

Zhimeng Guo, Teng Xiao+4 more|arXiv (Cornell University)|Apr 3, 2023

Advanced Graph Neural Networks7 citations

TL;DR

This survey provides a comprehensive overview of counterfactual learning on graph-structured data, categorizing existing methods into four problem types: fairness, explainability, link prediction, and real-world applications. It presents a unified framework, compiles key datasets, metrics, and open-source tools, and identifies critical challenges such as scalability, dynamic graphs, and unsupervised learning, positioning counterfactual learning as essential for interpretable, fair, and causally sound graph representation learning.

ABSTRACT

Graph-structured data are pervasive in the real-world such as social networks, molecular graphs and transaction networks. Graph neural networks (GNNs) have achieved great success in representation learning on graphs, facilitating various downstream tasks. However, GNNs have several drawbacks such as lacking interpretability, can easily inherit the bias of data and cannot model casual relations. Recently, counterfactual learning on graphs has shown promising results in alleviating these drawbacks. Various approaches have been proposed for counterfactual fairness, explainability, link prediction and other applications on graphs. To facilitate the development of this promising direction, in this survey, we categorize and comprehensively review papers on graph counterfactual learning. We divide existing methods into four categories based on problems studied. For each category, we provide background and motivating examples, a general framework summarizing existing works and a detailed review of these works. We point out promising future research directions at the intersection of graph-structured data, counterfactual learning, and real-world applications. To offer a comprehensive view of resources for future studies, we compile a collection of open-source implementations, public datasets, and commonly-used evaluation metrics. This survey aims to serve as a ``one-stop-shop'' for building a unified understanding of graph counterfactual learning categories and current resources. We also maintain a repository for papers and resources and will keep updating the repository https://github.com/TimeLovercc/Awesome-Graph-Causal-Learning.

Motivation & Objective

• To address the lack of interpretability and bias in graph neural networks (GNNs) by introducing counterfactual learning as a solution.
• To provide a unified framework for understanding diverse counterfactual learning approaches on graphs across multiple research problems.
• To compile and organize open-source implementations, public datasets, and evaluation metrics to serve as a centralized resource for researchers.
• To identify and highlight critical open challenges such as scalability, dynamic graph handling, and unsupervised counterfactual learning.
• To stimulate further research at the intersection of graph learning, causality, and real-world applications by outlining promising future directions.

Proposed method

• Categorizes existing graph counterfactual learning methods into four problem-based groups: counterfactual fairness, counterfactual explanation, counterfactual link prediction, and real-world applications.
• Proposes a general framework that unifies the problem formulation, data generation, model learning, and evaluation across all categories.
• Reviews representative methods in each category, emphasizing techniques for generating valid counterfactuals by perturbing graph structures (e.g., node/edge modifications).
• Emphasizes the use of causal inference principles to model 'what-if' scenarios, where the factual outcome is observed and the counterfactual outcome is inferred under intervention.
• Introduces evaluation protocols based on fidelity, plausibility, and sparsity to assess the quality of generated counterfactuals.
• Maintains a continuously updated GitHub repository with curated papers, code, datasets, and benchmarks to support reproducibility and community growth.

Experimental results

Research questions

• RQ1 How can counterfactual learning improve fairness in graph neural networks by mitigating bias from training data?
• RQ2 What are the key design principles and evaluation criteria for generating faithful and plausible counterfactual explanations on graph-structured data?
• RQ3 How can counterfactual reasoning enhance link prediction by modeling causal dependencies in graph structures?
• RQ4 What are the major challenges in scaling counterfactual learning to large-scale and dynamic graphs?
• RQ5 What opportunities exist for unsupervised or self-supervised counterfactual learning in low-resource graph settings?

Key findings

• Counterfactual learning effectively improves model interpretability and reduces bias in GNNs by generating 'what-if' scenarios that reveal causal dependencies.
• The survey identifies a lack of standardized evaluation metrics for counterfactual explanations on graphs, highlighting the need for metrics aligned with causal structure.
• Existing approaches struggle with scalability due to iterative counterfactual generation and retraining, especially on large graphs like social networks or knowledge graphs.
• Dynamic graph settings remain underexplored, with most methods assuming static graph structures, despite real-world graphs being inherently time-evolving.
• Unsupervised and self-supervised counterfactual learning on graphs is largely unexplored, despite the success of self-supervised pretraining in graph representation learning.
• The authors compile a comprehensive resource repository with open-source tools, datasets, and metrics, which is actively maintained to support ongoing research in graph counterfactual learning.