Contrastive unlearning via representation editing for graph neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph unlearning (GU) is crucial for privacy and data quality in graph neural networks (GNNs).
• Existing GU methods face challenges in balancing model utility with unlearning effectiveness.
• Difficulty in removing target data's influence and its propagated effects across the graph.

Purpose of the Study:

• To propose CURE, a novel graph unlearning framework using contrastive representation editing.
• To enhance the effectiveness and efficiency of graph unlearning.
• To improve privacy protection and data quality management in GNNs.

Main Methods:

• CURE employs an adaptive sample selection module to identify key nodes related to unlearning targets.
• Utilizes a contrastive unlearning strategy to decouple representations of forgotten nodes.
• Incorporates a personalized PageRank-based stability preservation module to maintain model utility.

Main Results:

• CURE achieves a favorable trade-off between model utility, unlearning efficiency, and effectiveness.
• Outperforms existing baseline methods in most experimental settings.
• Demonstrates superior privacy protection and robustness.

Conclusions:

• CURE offers an effective solution for graph unlearning challenges.
• The framework successfully removes target data influence while preserving model performance.
• CURE advances the state-of-the-art in privacy-preserving and robust graph learning.