FMP: Toward Fair Graph Message Passing against Topology Bias

Abstract

Despite recent advances in achieving fair representations and predictions through regularization, adversarial debiasing, and contrastive learning in graph neural networks (GNNs), the working mechanism (i.e., message passing) behind GNNs inducing unfairness issue remains unknown. In this work, we theoretically and experimentally demonstrate that representative aggregation in message passing schemes accumulates bias in node representation due to topology bias induced by graph topology. Thus, a Fair Message Passing (FMP) scheme is proposed to aggregate useful information from neighbors but minimize the effect of topology bias in a unified framework considering graph smoothness and fairness objectives. The proposed FMP is effective, transparent, and compatible with back-propagation training. An acceleration approach on gradient calculation is also adopted to improve algorithm efficiency. Experiments on node classification tasks demonstrate that the proposed FMP outperforms the state-ofthe-art baselines in effectively and efficiently mitigating bias on three real-world datasets.

Fairness in Graph Mining: A Survey

“…Regularization based on the network topology is proved to be effective in promoting group fairness. For instance, feature propagation is a common operation to model the dependency between neighboring nodes in graph mining [67], [131]. However, if the graph topology is biased, the propagated features also tend to be biased [40].…”

“…However, if the graph topology is biased, the propagated features also tend to be biased [40]. To tackle such problem, Jiang et al [67] achieved a less biased feature propagation through a fairness-aware regularization. Specifically, given two sensitive subgroups, the regularization is formally given as…”

“…To obtain a fair graph topology, edges are rewired in a greedy manner to maximally reduce the information unfairness score. Additionally, in GNNs, the information aggregation operation is found to introduce bias from the biased network topology to the learned node embeddings [40], [67], [81], [82], [91]. Dong et al [40] proposed to perform edge rewiring for fair node embedding learning.…”

“…German [9] group, couterfactual 1,000 21,742 27 gender (2) [3], [40], [173] NBA [38] group 403 10,621 96 country (2) [38], [67] Recividism [69] group, individual 18,876 311,870 18 race (2) [3], [40], [44], [104], [173] couterfactual Credit [168] group, couterfactual 30,000 1,421,858 13 age (2) [3], [40], [104], [173] in different real-world applications [23], [28], [134], [158], [169]. Although we have surveyed many fairness notions for graph mining, we need to admit that by no means are they complete, as other types of biases could also exist, depending on the needs of different real-world scenarios [160].…”

Subgroup Fairness in Graph-based Spam Detection

“…Generally, fairness issues on the graph are raised in two perspectives: (a) bias from the node feature [19,32] (b) topology bias coming from the edge connections [9,17,25,35]. In Figure 1, we give an example of a review graph where the edges describe that users leave reviews for products.…”