A Survey on Fairness for Machine Learning on Graphs

TL;DR

This survey addresses fairness in graph-based machine learning by organizing the problem around three fairness notions (group, individual, counterfactual) and three intervention points (pre-processing, in-processing, post-processing). It provides a structured review of metrics (structural, representation, task-specific), and method classes (pre-processing, fair node embeddings, end-to-end GNNs, post-processing) with a detailed catalog of datasets and benchmarks. Key contributions include a taxonomy of methods, a synthesis of synthetic and real-world benchmarks, and a discussion of open challenges such as non-IID data, homophily effects, multi-attribute fairness, and heterogeneous graphs. The work aims to guide researchers and practitioners toward principled, measurable, and scalable fair graph learning with practical pathways for evaluation and deployment.

Abstract

Nowadays, the analysis of complex phenomena modeled by graphs plays a crucial role in many real-world application domains where decisions can have a strong societal impact. However, numerous studies and papers have recently revealed that machine learning models could lead to potential disparate treatment between individuals and unfair outcomes. In that context, algorithmic contributions for graph mining are not spared by the problem of fairness and present some specific challenges related to the intrinsic nature of graphs: (1) graph data is non-IID, and this assumption may invalidate many existing studies in fair machine learning, (2) suited metric definitions to assess the different types of fairness with relational data and (3) algorithmic challenge on the difficulty of finding a good trade-off between model accuracy and fairness. This survey is the first one dedicated to fairness for relational data. It aims to present a comprehensive review of state-of-the-art techniques in fairness on graph mining and identify the open challenges and future trends. In particular, we start by presenting several sensible application domains and the associated graph mining tasks with a focus on edge prediction and node classification in the sequel. We also recall the different metrics proposed to evaluate potential bias at different levels of the graph mining process; then we provide a comprehensive overview of recent contributions in the domain of fair machine learning for graphs, that we classify into pre-processing, in-processing and post-processing models. We also propose to describe existing graph data, synthetic and real-world benchmarks. Finally, we present in detail five potential promising directions to advance research in studying algorithmic fairness on graphs.

Figures (8)

• Figure 1: Machine Learning on graphs. From the original graph (left) and eventually additional feature information (including the protected attribute), a mapping function (model) is learned that produces the matrix of node embeddings $Z$ (middle). These embeddings are then used to solve several types of graph problems (right): edge prediction (top), node classification (middle), or graph classification (bottom).
• Figure 2: Examples of networks with two different potential bias: job-candidate matching (a) and a social media network with political echo chambers (b) and in the presence of an influencer node (c).
• Figure 3: Illustration of the assortative mixing coefficient ($r$). Embeddings are computed with Node2Vec and a dimension of 8. The representation of node embeddings is done with TSNE.
• Figure 4: Illustration of the concept of Information Unfairness - taken from jalali2020. Node color indicates their sensitive group. (a) The left network has an IU of 0.047, while the right network has an IU of 0.1. It is clear that for the right network, very little information flows between red nodes. (b) Graphs with the same mixing coefficient, but different IU scores.
• Figure 5: Illustration of two GNN-based models for graph fairness: (a) MONETpalowitch2020monet which imposes independence (through orthogonality constraints) between the embedding spaces of features and topology; and (b) NIFTYagarwal2021towards that generates perturbated and counterfactual graphs to ensure stable node representation. Images are taken from original papers.

Theorems & Definitions (9)

• definition 1
• definition 2
• definition 3
• definition 4
• definition 5
• definition 6
• definition 7
• definition 8
• definition 9