Research map / Federated learning and privacy / Privacy attacks and secure learning
Reducing Item Discrepancy via Differentially Private Robust Embedding Alignment for Privacy-Preserving Cross Domain Recommendation
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 RidCDR, a privacy-preserving cross-domain recommender that aligns non-overlapping user/item embeddings across domains via differentially private projection plus robust sample-reweighted unbalanced optimal transport (DPRA).
Problem
Cross-domain recommendation when users and items are non-overlapping across domains and intra-domain interaction data cannot be shared due to privacy regulations, causing data sparsity and negative transfer from dissimilar items.
Approach
Combine a differentially private random projection with Gaussian noise on item embeddings and a sample-reweighting unbalanced optimal transport (SWAM + SROT) that down-weights dissimilar items to reduce domain discrepancy without shared users/items.
Key results
- On Amazon CD→Movie, RidCDR reaches HR@10 0.4088 and NDCG@10 0.2228, versus 0.3856 / 0.1902 for the strongest baseline MOTKD.
- On Amazon Video→Game, RidCDR reaches HR@10 0.4283 vs 0.4204 for MOTKD.
- Ablation: RidCDR-D (DP sliced OT) and RidCDR-U (entropic UOT) both underperform full RidCDR, showing the benefit of robust reweighting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon CD → Movie | HR@10: 0.3856 -> 0.4088 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon Movie → CD | HR@10: 0.4431 -> 0.4672 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon Video → Game | HR@10: 0.4204 -> 0.4283 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon CD → Book | HR@10: 0.3120 -> 0.3359 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon Book → CD | HR@10: 0.4253 -> 0.4384 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Amazon Game → Video | HR@10: 0.3455 -> 0.3627 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Douban Movie | HR@10: 0.3353 -> 0.3572 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Douban Book | HR@10: 0.3462 -> 0.3639 | Helped |
| use RidCDR rather than MOTKD | Cross-domain recommendation · Douban Music | HR@10: 0.3945 -> 0.4131 | Helped |
| add private-robust embedding alignment to the rating prediction module (ablation) | RidCDR · Amazon CD → Movie | HR@10: 0.3067 -> 0.4088 | Helped |
| use DPRA rather than DP-Sliced OT for embedding alignment (ablation) | RidCDR · Amazon CD → Movie | HR@10: 0.3904 -> 0.4088 | Helped |
| use DPRA rather than entropic UOT for embedding alignment (ablation) | RidCDR · Amazon CD → Movie | HR@10: 0.3991 -> 0.4088 | Helped |
Limitations
- Privacy guarantee relies on a single random projection plus Gaussian noise with a moderate budget epsilon=4, and privacy analysis is borrowed from prior work
- Gains over strongest baselines are modest on several benchmarks
- Experiments limited to Amazon and Douban datasets with binary rating setup
Experimental setup
Datasets: Amazon Movie Amazon Book Amazon CDs Amazon Videos Amazon Games Douban Movie Douban Book Douban Music
Benchmarks: Amazon CD-Movie Amazon Video-Game Amazon CD-Book Amazon Game-Video Douban Movie-Book-Music
Metrics: HR@10 NDCG@10 HR@20 NDCG@20
Baselines: NeuMF DeepCoNN NARRE DDTCDR Rec-DAN CGN TDAR CDRIB DisAlign CFAA GWCDR DURation
Builds on: Sliced optimal transport DP (Rakotomamonjy & Liva, 2021) unbalanced optimal transport MOTKD BERT
Closest papers by meaning
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.87)
- Towards Unbiased Information Extraction and Adaptation in Cross-Domain Recommendation (AAAI 2025 · similarity 0.86)
- Learning Accurate and Bidirectional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation (AAAI 2024 · similarity 0.85)
- Domain-Level Disentanglement Framework Based on Information Enhancement for Cross-Domain Cold-Start Recommendation (AAAI 2025 · similarity 0.85)
- Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation (AAAI 2024 · similarity 0.84)
- RRL: Recommendation Reverse Learning (AAAI 2024 · similarity 0.84)
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.