Research map / Vision-language and multimodal / Cross-modal learning
Multi-Modal Recommendation Unlearning for Legal, Licensing, and Modality Constraints
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 MMRecUn, a reverse-BPR-based unlearning method for multi-modal graph recommenders that removes user, item, or user-item interaction data while retaining a retain-set objective, evaluated on Amazon Baby, Sports and Clothing.
Problem
Unlearning in multi-modal recommender systems (MMRS), especially item-side removal (licensing, legal, outdated preferences), which prior RS unlearning methods (matrix-based, sharded exact methods) do not handle well.
Approach
Fine-tunes a trained MGCN model with a reverse BPR loss on the forget set, balanced against forward BPR plus negated contrastive and L2 terms on the retain set, to forget marked interactions without sharding or retraining from scratch.
Key results
- Unlearning 5% users, MMRecUn improves test recall over AmUn by up to 46.93% and validation recall by up to 49.85%, approaching Gold.
- On item unlearning, MMRecUn's forget-set recall is 10.83% and 8.45% higher than AmUn on two datasets, with comparable test recall.
- MMRecUn unlearning takes 169.41s vs 225.73s for Gold on Baby (about 1.3x faster).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MMRecUn instead of AmUn for user unlearning | MGCN · Baby, Sports, and Clothing user unlearning 5% | test Recall@20: 26.69%, 46.93%, and 36.13% | Helped |
| Apply MMRecUn instead of AmUn for user unlearning | MGCN · Baby, Sports, and Clothing user unlearning 5% | validation Recall@20: 29.6%, 49.85%, and 34.75% | Helped |
| Apply MMRecUn instead of AmUn for item unlearning | MGCN · Baby, Sports, and Clothing item unlearning 5% | forget-set Recall: 10.83%, 8.45%, and 3.43% | Mixed |
| Apply MMRecUn instead of AmUn for simultaneous user-item unlearning | MGCN · User-centric metrics across three datasets | Recall: 2.89%, 1.83% and 0% | No clear effect |
| Use MMRecUn instead of retraining the Gold model | MGCN · Baby and Clothing unlearning | unlearning time: up to 1.3×; 25% | Helped |
| Set α to 0.1–0.2 (ablation) | MMRecUn · Baby user unlearning 5% | validation Recall@20: 0.0893 | Helped |
| Increase α to 0.3–0.9 (ablation) | MMRecUn · Baby user unlearning 5% | retained-set recall | Hurt |
| Set α below 0.01 (ablation) | MMRecUn · Baby user unlearning 5% | forget-set recall | Hurt |
Surprising findings
- The reported MMRecUn gain over AmUn on forget recall for clothing users is slightly negative (-2.08%), and MMRecUn falls back on sports items (-3.43%).
- Baseline Gold has near-zero forget-set recall on user-item unlearning in some settings, and AmUn is a repurposed image-classification method rather than an RS-specific one.
Limitations
- Only one base model (MGCN) and three Amazon categories are evaluated; sharded and approximate RS unlearning baselines (RecEraser, UltraRE, Rrl, Ifru, Scif) were not compared due to unavailable code or scope.
- Speedup over Gold is modest (about 25% time saved), and the gains are reported relative to a repurposed image-unlearning baseline.
- Hyperparameter alpha must be tuned per task; the paper's Bayesian and information-bound propositions are informal.
Experimental setup
Datasets: Amazon Baby Amazon Sports Amazon Clothing
Benchmarks: Amazon Baby Amazon Sports Amazon Clothing
Metrics: Recall@K Precision@K NDCG@K MAP unlearning time
Baselines: Gold model AmUn MGCN
Builds on: MGCN BPR amnesiac unlearning SISA
Closest papers by meaning
- CURE4Rec: A Benchmark for Recommendation Unlearning with Deeper Influence (NeurIPS 2024 · similarity 0.87)
- RRL: Recommendation Reverse Learning (AAAI 2024 · similarity 0.85)
- Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation (NeurIPS 2024 · similarity 0.83)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.83)
- MTSTRec: Multimodal Time-Aligned Shared Token Recommender (ICML 2025 · similarity 0.83)
- RecFlow: An Industrial Full Flow Recommendation Dataset (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.