Recommendation Unlearning with Dynamic Sampling and Interest Boundary Perception

Machine unlearning has emerged as a pivotal technique for privacy-preserving recommendation systems, enabling users to safeguard their privacy. Existing recommender systems utilizing unlearning typically undergo static sampling of partitions from the remaining data, excluding those samples designated for forgetting. However, two essential drawbacks still remain:1) static sampling may lead to unnecessary memory consumption and slower convergence of the model; and 2) the loss of boundaries by different sub-partitions during aggregation produces sub-optimal prediction performance. To address these challenges, we present a DI-RU model. In the DI-RU framework, we tackle the first limitation by introducing a dynamic negative sampling mechanism. This mechanism selects negative samples from the ranked list produced by the current predictive model, facilitating rapid iterative model updates. To counteract the second limitation, we implement a boundary-penalization strategy. This strategy incorporates two auxiliary scores for each user, which represent the boundaries of their preferences and dislikes during the aggregation phase. Samples that traverse these boundaries are subjected to individual penalties. The results of our experiments on the three public datasets indicate that the proposed DI-RU outperforms several other baseline systems, we present a case study through a real-world example showcasing the usefulness of DI-RU for generating explanations.