Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Unlearning with Efficient Partial Retraining

Jiahao Zhang, Lin Wang, Shijie Wang, Wenqi Fan

TL;DR

The paper tackles unlearning in graph neural networks by addressing the limitations of existing retraining-based methods, which often degrade sub-model utility when partitions erode structural/semantic information. It introduces GraphRevoker, a two-part framework: (i) graph property-aware sharding that uses a differentiable partition network to create $S$ subgraphs while optimizing unlearning time, edge cuts via $Ncut$, and label distribution via entropy; and (ii) graph contrastive sub-model aggregation that ensembles sub-models with attentive fusion and a local-global contrastive loss together with a local-local reconstruction loss. The aggregation module, trained on a small labeled subset, yields strong predictive performance with efficient unlearning, and the whole approach fits within the retraining-based paradigm to guarantee exact removal of undesirable data. Extensive experiments on four real-world datasets show GraphRevoker outperforms existing efficient unlearning methods in both model utility and unlearning time, with ablations confirming the necessity of the partitioning and aggregation components. Overall, GraphRevoker offers a scalable, exact, and utility-preserving solution for graph unlearning in GNNs, enabling safe removal of undesirable data in large-scale graphs.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success in various real-world applications. However, GNNs may be trained on undesirable graph data, which can degrade their performance and reliability. To enable trained GNNs to efficiently unlearn unwanted data, a desirable solution is retraining-based graph unlearning, which partitions the training graph into subgraphs and trains sub-models on them, allowing fast unlearning through partial retraining. However, the graph partition process causes information loss in the training graph, resulting in the low model utility of sub-GNN models. In this paper, we propose GraphRevoker, a novel graph unlearning framework that better maintains the model utility of unlearnable GNNs. Specifically, we preserve the graph property with graph property-aware sharding and effectively aggregate the sub-GNN models for prediction with graph contrastive sub-model aggregation. We conduct extensive experiments to demonstrate the superiority of our proposed approach.

Graph Unlearning with Efficient Partial Retraining

TL;DR

The paper tackles unlearning in graph neural networks by addressing the limitations of existing retraining-based methods, which often degrade sub-model utility when partitions erode structural/semantic information. It introduces GraphRevoker, a two-part framework: (i) graph property-aware sharding that uses a differentiable partition network to create subgraphs while optimizing unlearning time, edge cuts via , and label distribution via entropy; and (ii) graph contrastive sub-model aggregation that ensembles sub-models with attentive fusion and a local-global contrastive loss together with a local-local reconstruction loss. The aggregation module, trained on a small labeled subset, yields strong predictive performance with efficient unlearning, and the whole approach fits within the retraining-based paradigm to guarantee exact removal of undesirable data. Extensive experiments on four real-world datasets show GraphRevoker outperforms existing efficient unlearning methods in both model utility and unlearning time, with ablations confirming the necessity of the partitioning and aggregation components. Overall, GraphRevoker offers a scalable, exact, and utility-preserving solution for graph unlearning in GNNs, enabling safe removal of undesirable data in large-scale graphs.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success in various real-world applications. However, GNNs may be trained on undesirable graph data, which can degrade their performance and reliability. To enable trained GNNs to efficiently unlearn unwanted data, a desirable solution is retraining-based graph unlearning, which partitions the training graph into subgraphs and trains sub-models on them, allowing fast unlearning through partial retraining. However, the graph partition process causes information loss in the training graph, resulting in the low model utility of sub-GNN models. In this paper, we propose GraphRevoker, a novel graph unlearning framework that better maintains the model utility of unlearnable GNNs. Specifically, we preserve the graph property with graph property-aware sharding and effectively aggregate the sub-GNN models for prediction with graph contrastive sub-model aggregation. We conduct extensive experiments to demonstrate the superiority of our proposed approach.
Paper Structure (25 sections, 18 equations, 3 figures, 6 tables)

This paper contains 25 sections, 18 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Related Work
  4. Proposed Method
  5. Graph Property-Aware Sharding
  6. Graph Contrastive Sub-model Aggregation
  7. Experiment
  8. Conclusion
  9. Supplementary Material
  10. Evaluation Settings
  11. Dataset Statistics.
  12. General Settings.
  13. Model Utility Settings.
  14. Unlearning Efficiency Settings.
  15. Additional Experiments
  16. ...and 10 more sections

Figures (3)

  • Figure 1: An example of graph unlearning in a social network.
  • Figure 2: Illustration of the proposed framework.
  • Figure 3: The unlearning power of our proposed framework. This figure shows the F1-score (%) on Cora and Citeseer of the original model, the model affected by low-quality data, and the model unlearned the low-quality data.