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.

Original post →

More from Safety

Safety channel →