Selective and collaborative influence function for efficient recommendation unlearning

Recent regulations concerning the Right to be Forgotten have greatly influenced operation of recommender systems, because users now right withdraw their private data. Besides simply deleting target data in database, unlearning associated lineage e.g., learned personal features and preferences model, is also necessary for withdrawal. Existing methods are mainly devised generalized machine learning models classification tasks. In this paper, we first identify two main disadvantages directly applying existing context recommendation, i.e., (i) unsatisfactory efficiency large-scale recommendation (ii) destruction collaboration across items. To tackle above issues, propose a highly efficient method based on Selective Collaborative Influence Function (SCIF). Our proposed can avoid any kind retraining which computationally prohibitive further enhance by selectively updating user embedding (iii) preserve remaining Furthermore, order evaluate completeness, define Membership Inference Oracle (MIO) that verifies whether unlearned points were part model’s training set, thereby determining if point was completely unlearned. Extensive experiments benchmark datasets demonstrate our not only efficiency, but achieve adequate completeness. More importantly, outperforms State-Of-The-Art (SOTA) regarding comprehensive metrics.