An Efficient Graph-Based Recommendation Unlearning via Knowledge Distillation

Recommender systems rely on user--item interactions to provide personalized services, yet privacy regulations such as the \textit{Right to be Forgotten} necessitate effective recommendation unlearning. Graph Collaborative Filtering (GCF) models pose a particular challenge because message passing distributes the influence of an interaction across multi-hop neighbors. Consequently, removing forgotten interactions can induce unintended changes in other interactions, leading to incomplete forgetting or unnecessary modification of retained knowledge. We propose RecRemover, a model-agnostic selective framework for graph recommendation unlearning. RecRemover introduces \textit{Topology-Aware Affected Interaction Sampling (TAIS)}, which uses predicted score shifts induced by removing the forget interactions to identify affected forget and retain interactions. Based on these subsets, RecRemover jointly optimizes a student model using an unlearning objective that suppresses the influence of affected forget interactions and a learning objective that preserves the preference behavior of the original model on affected retain interactions. By focusing optimization on propagation-affected interactions, RecRemover avoids indiscriminate updates while addressing the effects of interaction deletion. We evaluate RecRemover on three real-world datasets using two GCF backbones and compare it with retraining and state-of-the-art recommendation unlearning methods. Results show that RecRemover achieves high recommendation utility, effective unlearning, strong ranking consistency with the retrained model, and improved unlearning efficiency.