<ul><li>The right to be forgotten regulations have led to advancements in certified unlearning strategies for graph data to comply with data removal requests from machine learning models.</li><li>A training-free unlearning procedure is developed for pre-trained graph ML models to mitigate biases, offering certifiable bias mitigation through a single-step Newton update on model weights.</li><li>The approach provides a computationally lightweight alternative to existing fairness enhancement methods, with quantifiable performance guarantees.</li><li>Experimental results demonstrate the efficacy of the developed unlearning strategies in mitigating biases while maintaining minimal impact on node classification accuracy, showcasing favorable utility-complexity trade-offs compared to retraining models from scratch.</li></ul>

Unlearning Algorithmic Biases over Graphs

Discover more