Research map / Federated learning and privacy / Privacy attacks and secure learning
RRL: Recommendation Reverse Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Recommendation Reverse Learning (RRL), which fine-tunes a trained recommender with a reversed BPR objective regularized by a Fisher-information collaborative similarity term to forget marked user interactions without splitting data.
Problem
Right-to-be-forgotten unlearning for recommender systems: existing retrain-based (shard) methods like SISA and RecEraser are efficient but degrade recommendation accuracy because splitting data breaks collaborative information.
Approach
Adapts reverse learning to recommendation by defining a reversed personalized ranking (RPR) objective that pushes marked interactions' scores below non-interacted items, and adds a Fisher-information-based collaborative similarity regularizer so updates to embeddings also account for related users and items; theoretically linked to retraining via a Bayesian view.
Key results
- On Gowalla with MF-BPR, RRL matches Retrain (Recall 0.144 vs 0.143) while reducing unlearning time to 3.5s versus 12841s for Retrain.
- SISA and RecEraser degrade recommendation on sparse Gowalla and Yelp2018 (e.g., SISA Recall 0.051 vs 0.144 on Gowalla MF-BPR).
- Shilling attack target-item recommendations drop to 0 on MF-BPR after RRL unlearning, and by roughly 80% on LightGCN.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RRL instead of retraining | MF-BPR · Gowalla | Recall@20: 0.143 -> 0.144 | No clear effect |
| Use RRL instead of retraining | LightGCN · Yelp2018 | Recall@20: 0.063 -> 0.063 | No clear effect |
| Use RRL instead of retraining | LightGCN · ML-1M | Recall@20: 0.243 -> 0.250 | Helped |
| Add the CS regularizer (ablation) | RRL with MF-BPR · Popularity Attack | Recall@20: 0.134 -> 0.144 | Helped |
| Add the CS regularizer (ablation) | RRL with MF-BPR · Popularity Attack | #Rec: 1 -> 3 | Hurt |
| Add the CS regularizer (ablation) | RRL with LightGCN · Popularity Attack | Recall@20: 0.174 -> 0.170 | Hurt |
| Add the CS regularizer (ablation) | RRL with LightGCN · Popularity Attack | #Rec: 716 -> 126 | Helped |
| Add the CS regularizer (ablation) | RRL with MF-BPR · Bilevel Attack | Recall@20: 0.135 -> 0.144 | Helped |
| Add the CS regularizer (ablation) | RRL with MF-BPR · Bilevel Attack | #Rec: 98 -> 1 | Helped |
| Add the CS regularizer (ablation) | RRL with LightGCN · Bilevel Attack | Recall@20: 0.170 -> 0.172 | Helped |
| Add the CS regularizer (ablation) | RRL with LightGCN · Bilevel Attack | #Rec: 786 -> 299 | Helped |
Surprising findings
- On the denser ML-1M dataset, SISA and RecEraser outperform Retrain and RRL in recommendation performance, since sharded aggregation helps there.
Limitations
- Performance degradation of 10-30% reported in the paper for retrain-based methods is the motivation; RRL's CS regularizer requires FIM computation offline after full training.
- Verification relies on shilling attacks rather than membership inference, since no efficient MIA exists for recommenders per the authors.
- Evaluated only on two embedding-based recommenders and three datasets with a single unlearning setting of 100 users.
Experimental setup
Datasets: Gowalla Yelp2018 ML-1M
Benchmarks: Gowalla Yelp2018 ML-1M
Metrics: Recall@20 NDCG@20 running time (RT) #Rec of target item under shilling attacks
Baselines: Retrain SISA RecEraser
Builds on: BPR SISA RecEraser Reverse learning up-weighted loss (Graves et al.)
Closest papers by meaning
- CURE4Rec: A Benchmark for Recommendation Unlearning with Deeper Influence (NeurIPS 2024 · similarity 0.88)
- Multi-Modal Recommendation Unlearning for Legal, Licensing, and Modality Constraints (AAAI 2025 · similarity 0.85)
- Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation (ICML 2024 · similarity 0.84)
- PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation (NeurIPS 2024 · similarity 0.83)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.83)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.