Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing

TL;DR

This work addresses bias introduced by graph unlearning, proposing FGU, a shard-based, fairness-aware framework that jointly enforces shard-level and global fairness through bi-level debiasing. By partitioning the graph into shards, retraining locally with a DP-based fairness regularizer, and globally aligning shard predictions, FGU preserves privacy while reducing disparities across sensitive groups. Extensive experiments on six datasets show that FGU outperforms standard and approximate graph unlearning baselines in fairness with competitive accuracy, and maintains robustness across varying data deletions and distributions. The results highlight the practical potential of fair graph unlearning for privacy-sensitive applications in social networks, recommendations, and beyond.

Abstract

Graph unlearning is a crucial approach for protecting user privacy by erasing the influence of user data on trained graph models. Recent developments in graph unlearning methods have primarily focused on maintaining model prediction performance while removing user information. However, we have observed that when user information is deleted from the model, the prediction distribution across different sensitive groups often changes. Furthermore, graph models are shown to be prone to amplifying biases, making the study of fairness in graph unlearning particularly important. This raises the question: Does graph unlearning actually introduce bias? Our findings indicate that the predictions of post-unlearning models become highly correlated with sensitive attributes, confirming the introduction of bias in the graph unlearning process. To address this issue, we propose a fair graph unlearning method, FGU. To guarantee privacy, FGU trains shard models on partitioned subgraphs, unlearns the requested data from the corresponding subgraphs, and retrains the shard models on the modified subgraphs. To ensure fairness, FGU employs a bi-level debiasing process: it first enables shard-level fairness by incorporating a fairness regularizer in the shard model retraining, and then achieves global-level fairness by aligning all shard models to minimize global disparity. Our experiments demonstrate that FGU achieves superior fairness while maintaining privacy and accuracy. Additionally, FGU is robust to diverse unlearning requests, ensuring fairness and utility performance across various data distributions.

Figures (5)

• Figure 1: Fairness performance of a two-layer GCNs before (left) and after (right) applying and . $10\%$ of the nodes and $10\%$ of the edges were randomly sampled from the unprivileged sensitive group for unlearning.
• Figure 2: An illustration of : shard debiasing achieves equitable predictions within each shard model, and global alignment reduces disparities among sensitive groups across shard models.
• Figure 3: Accuracy of graph unlearning methods on three datasets, under $10\%$ of node unlearning ratio and different edge unlearning ratios.
• Figure 4: Accuracy and $\triangle_{DP}$ trade-off comparison on three datasets and five fairness-aware GNNs. Results located in the bottom-right corner are preferable.
• Figure 5: Accuracy and $\triangle_{DP}$ trade-off on German, Credit, and Pokec-n. Results located in the bottom-right corner are preferable.