Graph unlearning aims to remove the influence of specific training data from a trained graph neural network without retraining from scratch. Most existing approaches focus on removing nodes or features, but the problem of forgetting edges has received less attention.
The paper introduces Lethe, a link inference attack that evaluates edge unlearning methods. By attempting to infer whether a particular edge was part of the training data, Lethe provides a way to measure how effectively an unlearning method erases edge information.
The abstract does not report specific experimental results, so the effectiveness of Lethe on different models remains to be seen. Still, it highlights a gap in the current evaluation toolkit for edge unlearning.