SURF: subtractive unlearning for recommenders at 2% of full-retraining cost
_reachsumit · x · 2026-09-17
Filippo Betello et al. propose SURF, a lightweight approximate machine-unlearning framework for sequential recommender systems, driven by GDPR-style privacy compliance.
It works in three steps: locate the to-be-forgotten item's neighborhood in embedding space, train an auxiliary model on that local subset, then subtract the auxiliary scores from the original model at inference. Versus baselines requiring full retraining or ignoring sequential behavior, SURF matches full-retraining unlearning quality across 7 datasets and 5 baselines, improving NDCG@20 by up to 32% while using only 2% of the retraining time budget. Code is open-sourced.
More from Safety
- VraserX backs independent AI capability evals, slams incumbent-style 'safety' — VraserX · 2026-09-17
- Leak claims iPhone 18 Pro signs photos at sensor level with quantum-safe crypto to beat AI fakes — jackedAJ · 2026-09-17
- iPhone 18 Pro rumored to sign photos at sensor level with quantum-safe cryptography — jackedAJ · 2026-09-17
- OpenAI reveals 'concerning' AI behaviour cases, promises new disclosure plan — kiyomoris · 2026-09-17
- AI scientist Toby Walsh: 'kill us all' scenario fantastical, real dangers remain — TobyWalsh · 2026-09-17
- Ethics of Simulation: Extend to Others the Complexity You'd Want for Yourself — basedjensen · 2026-09-17