Re-understanding Graph Unlearning through Memorization

TL;DR

This work reframes graph unlearning through the lens of GNN memorization, introducing MGU, a memorization-guided framework that (i) provides practical, test-data-free difficulty assessment, (ii) adapts unlearning objectives via a margin-based forgetting and distillation-based preservation strategy, and (iii) establishes a ToU-based comprehensive evaluation protocol. Empirically, MGU achieves state-of-the-art forgetting quality and computational efficiency across ten real-world graphs and multiple GNN backbones, outperforming baselines especially on hard tasks. The approach also demonstrates strong robustness to adversarial perturbations and broad strategy generalizability when integrated with existing GU methods. By linking memorization to unlearning difficulty and using difficulty-aware evaluation, the work offers a practical, scalable path toward reliable, real-world graph unlearning.

Abstract

Graph unlearning (GU), which removes nodes, edges, or features from trained graph neural networks (GNNs), is crucial in Web applications where graph data may contain sensitive, mislabeled, or malicious information. However, existing GU methods lack a clear understanding of the key factors that determine unlearning effectiveness, leading to three fundamental limitations: (1) impractical and inaccurate GU difficulty assessment due to test-access requirements and invalid assumptions, (2) ineffectiveness on hard-to-unlearn tasks, and (3) misaligned evaluation protocols that overemphasize easy tasks and fail to capture true forgetting capability. To address these issues, we establish GNN memorization as a new perspective for understanding graph unlearning and propose MGU, a Memorization-guided Graph Unlearning framework. MGU achieves three key advances: it provides accurate and practical difficulty assessment across different GU tasks, develops an adaptive strategy that dynamically adjusts unlearning objectives based on difficulty levels, and establishes a comprehensive evaluation protocol that aligns with practical requirements. Extensive experiments on ten real-world graphs demonstrate that MGU consistently outperforms state-of-the-art baselines in forgetting quality, computational efficiency, and utility preservation.

Figures (8)

• Figure 1: Unlearning performance on the 10% easiest and 10% hardest nodes in the Cora dataset. The red dashed line denotes the gold-standard performance obtained by retraining.
• Figure 2: Long-tailed distribution of node unlearning difficulty (via Eq. (\ref{['eq2']})). Higher scores indicate harder cases.
• Figure 3: ToU performance of GU methods for node unlearning on Cora with GAT backbone ($N=5$).
• Figure 4: Impact on generalization after removing nodes at varying difficulty levels.
• Figure 5: Adaptive Unlearning Strategy.