Obliviate: Efficient Unlearning Framework for Recommender Systems
A new machine unlearning framework called Obliviate addresses data privacy regulations for recommender systems. It removes user interaction data and its downstream influence from trained models without full retraining. The two-stage approach uses a Low-Rank Unlearning Adapter (LUA) with a lightweight Hessian proxy for curvature-aware unlearning. Existing methods suffer from incomplete unlearning, performance degradation, and high computational cost. Obliviate achieves high unlearning completeness while maintaining utility. The paper is published on arXiv under ID 2607.22665.
Key facts
- Obliviate is a two-stage unlearning framework for recommender systems.
- It uses a Low-Rank Unlearning Adapter (LUA) with a Hessian proxy.
- The goal is to remove user interaction data and its influence without full retraining.
- Existing approaches have limitations in completeness, performance, and computational cost.
- Obliviate achieves high unlearning completeness while preserving recommendation quality.
- The paper is available on arXiv with ID 2607.22665.
- Machine unlearning is critical for data privacy regulations.
- The framework is efficient and curvature-aware.
Entities
Institutions
- arXiv