Recommender systems require on-demand unlearning of user-item interactions to meet privacy regulations without sacrificing recommendation quality. Existing approaches either retrain large models at prohibitive cost or apply local parameter tweaks that misestimate non-linear effects, resulting in degraded accuracy, poor scalability for bulk deletions, and no formal privacy auditing. We present ReCUR, a Recommendation Contrastive Unlearning framework with Influence Estimation that addresses these weaknesses. ReCUR first applies a contrastive push-pull loss to push forgotten interaction embeddings away and pull retained ones back toward their original anchors. This not only subtracts deleted interactions’ contributions from the model, but yields fine-grained forgetting control, minimizes collateral drift, and recovers lost utility without full retraining. To compensates for any lost utility and promote fairness, ReCUR introduces a re-ranking mechanism with group-specific promotion weights that rebalance recommendations for diverse and niche user segments. Across four real-world datasets, ReCUR matches or exceeds retrain-from-scratch accuracy, retaining over 95% utility even after deleting up to 20% of interactions, while achieving up to an order-of-magnitude speed-up. Membership-inference audits confirm ReCUR’s unlearned models are indistinguishable from fresh retrains, offering strong empirical privacy guarantees. Our further analysis reveals intuitive trade-offs between forgetting strength and utility, and shows that promoting under-represented users enhances overall accuracy.
Yang et al. (Tue,) studied this question.