Machine Unlearning Scheme for Recommendation System Based on Gradient Attribution

Addressing the current data privacy leakage challenges faced by BERT-based sequential recommendation systems, this study innovatively proposes a gradient attribution-based machine unlearning scheme for recommendation systems. The scheme first leverages gradient attribution techniques to precisely evaluate the contribution of each intermediate neuron to the prediction of user behavior information. Based on these evaluation results, we adopt a filtering strategy designed to identify and remove redundant neurons while retaining those "privacy neurons" closely associated with specific sensitive user behavior information. Subsequently, the activation values of these identified privacy neurons are set to zero, thereby completely eliminating their contribution to model prediction. This effectively achieves the goal of erasing specific categories of user behavior information from the recommendation system model. Compared to traditional model retraining methods, this scheme offers a significant advantage in unlearning speed and has a minimal impact on the overall performance of the recommendation system, providing an efficient and practical new approach to addressing data privacy issues in sequential recommendation systems.