Counterfactual Learning on Graphs: A Survey (2304.01391v3)

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.

• The paper presents a detailed taxonomy of counterfactual methods on graphs, highlighting adversarial debiasing and perturbative techniques in GNN training.
• The paper examines counterfactual explanation strategies that enhance model transparency by identifying minimal input changes leading to outcome shifts.
• The paper analyzes counterfactual link prediction approaches that mitigate bias in graph connectivity, offering insights for practical applications.

An Essay on "Counterfactual Learning on Graphs: A Survey"

The paper "Counterfactual Learning on Graphs: A Survey" provides a comprehensive review on the burgeoning field of counterfactual learning applied to graph-structured data. As graphs inherently represent complex interdependencies in various domains such as social networks, molecular biology, and knowledge graphs, there is an exigent need to understand and mitigate inherent biases, improve interpretability, and enhance predictive accuracy through counterfactual reasoning. The survey categorizes existing works into four main domains: counterfactual fairness, counterfactual explanation, counterfactual link prediction, and domain-specific applications, offering a detailed taxonomy and summarizing promising future directions.

In the field of counterfactual fairness on graphs, the surveyed methods predominantly focus on the biases ingrained in node features and graph structures, which can propagate through Graph Neural Networks (GNNs) to yield discriminatory outcomes. The paper meticulously categorizes existing fairness techniques into adversarial debiasing, fairness constraint strategies, and counterfactual-based methods. A recurring theme is the employment of perturbative and generative techniques to construct counterfactual scenarios, thereby facilitating fairness. Specifically, the discussion extends to frameworks such as GEAR and NIFTY, which utilize neighbor perturbations and triplet-based objectives to bolster fairness. A critical observation is the predominant reliance on in-processing strategies that integrate fairness during GNN training, contrasting with the less explored pre- and post-processing phases.

Moving to counterfactual explanations, the survey highlights the importance of distilling actionable insights from GNNs by identifying minimal changes leading to different predictions. This nuanced understanding helps filter spurious correlations and enhances model transparency in high-stakes applications. The paper reviews significant approaches like CF-GNNExplainer and CF$^2$, which utilize graph perturbations and factual counterfactual regularization to derive necessary and sufficient explanations. An insightful framework is provided, focusing on the input and output levels of counterfactual reasoning, which aids in aligning the explanation methodologies across diverse graph architectures.

In counterfactual link prediction and recommendation, there is an exploration of novel causal perspectives to unearth root causes of link formations. CFLP, a representative paper, formulates the link prediction problem by considering both factual and counterfactual link distributions, offering a counterfactual lens to identify and mitigate spurious linkages. In the domain of recommendations, counterfactual methodology aids in tackling long-standing issues like clickbait and geographical bias by modeling direct and indirect effects of confounding factors.

The survey underscores the practical relevance of graph counterfactual learning in applications extending to physical systems, medical diagnostics, and molecular chemistry. Notable mentions are methodologies employing counterfactual learning to predict object trajectories, elucidate medical predictions, and decipher molecular interactions, showing vast potential in translating theoretical models into real-world impact.

Looking ahead, the paper identifies promising areas such as developing benchmark datasets tailor-made for counterfactual tasks, creating dedicated evaluation metrics, and venturing into unsupervised realms of counterfactual learning. Moreover, scalability and dynamic graph settings remain as pressing challenges, demanding scalable solutions that intelligently handle large-scale and temporally evolving graph data.

Overall, this survey extensively maps the landscape of counterfactual learning on graphs, providing valuable insights into the methodologies and applications while paving the path for future research avenues. The synthesis of counterfactual reasoning with graph learning paradigms offers an enriched toolkit for scientists aiming to enhance fairness, interpretability, and generalizability in graph-based models.