Abstract
With increasing data privacy regulations, machine unlearning is becoming crucial in recommendation systems. The goal of this work is to remove requested interaction data and their downstream influence from trained models while preserving recommendation quality and avoiding the substantial computational cost of full retraining. Existing approaches have several limitations, including limited unlearning completeness and degradation in recommendation performance, alongside significant computational overhead.
To address this, we propose Obliviate, an efficient two-stage unlearning framework for recommender systems that achieves high unlearning completeness while maintaining good utility. In the first stage, we introduce a Low-Rank Unlearning Adapter (LUA), which employs a lightweight Hessian proxy to enable curvature-aware and efficient unlearning through localized low-rank adapters instead of full parameters. In the second stage, we propose Locality-Aware Calibration (LAC), a lightweight refinement stage that updates only the adapter parameters to improve performance by enforcing unlearning via ranking-based objectives while preserving utility through knowledge distillation.
Extensive empirical evaluations demonstrate that Obliviate achieves a high level of forgetting with minimal loss in recommendation quality at significantly reduced computational costs, offering a practical and scalable solution for large-scale recommender systems.
Blogger's Review: The Obliviate framework effectively addresses the unlearning challenge in recommendation systems, showcasing significant potential in real-world applications while safeguarding user privacy. Its design leveraging low-rank adapters and locality-aware calibration provides a fresh perspective for optimizing future recommendation systems.