Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Recommendation with Explicitly Encoding Item Bias
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 FREIB, a platform-level federated recommender that explicitly encodes item bias with a separate encoder, uses global-model distillation and rating-level prototype alignment to handle label skew.
Problem
Platform-level federated recommendation suffers from severe label skew, where differing rating habits across platforms (rating bias) and item intrinsic quality (item bias) confuse item embedding learning and hurt generalization.
Approach
Adds an explicit item bias embedding and score function added to the NCF prediction, distills knowledge from a frozen global model to guide local training, and aligns local item-bias feature prototypes per rating level to global prototypes.
Key results
- FREIB achieves MAE 0.7369 and RMSE 0.9395 on MovieLens-100K (beta=1), versus 0.8650 MAE for the best prior baseline PFedRec on MovieLens-1M (beta=1) and 0.9191 on MovieLens-100K for PFedRec's counterpart row.
- On MovieLens-1M (beta=1) FREIB reaches MAE 0.7275, RMSE 0.9241, outperforming all federated baselines.
- Ablation on MovieLens-100K shows removing item bias encoder and prototype alignment together yields the largest degradation (MAE 0.9449).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FREIB instead of PFedRec | Platform-level federated recommendation · MovieLens-100K (β=1) | MAE: 1.1708 -> 0.7369 | Helped |
| use FREIB instead of PFedRec | Platform-level federated recommendation · MovieLens-1M (β=1) | MAE: 0.8650 -> 0.7275 | Helped |
| use FREIB instead of PFedRec | Platform-level federated recommendation · Beauty (β=1) | MAE: 1.0969 -> 0.9579 | Helped |
| use FREIB instead of PFedRec | Platform-level federated recommendation · MovieLens-100K (β=0.5) | MAE: 0.9191 -> 0.7926 | Helped |
| use FREIB instead of PFedRec | Platform-level federated recommendation · MovieLens-1M (β=0.5) | MAE: 0.8431 -> 0.7454 | Helped |
| use FREIB instead of PFedRec | Platform-level federated recommendation · Beauty (β=0.5) | MAE: 1.2962 -> 1.1635 | Helped |
| add explicit item bias encoder (ablation) | FREIB · MovieLens-100K and MovieLens-1M (β=1) | MAE: 0.9449 -> 0.7369; 0.9272 -> 0.7275 | Helped |
| add global knowledge guidance (ablation) | FREIB · MovieLens-100K and MovieLens-1M (β=1) | MAE: 0.7767 -> 0.7369; 0.7788 -> 0.7275 | Helped |
| add feature prototype alignment (ablation) | FREIB · MovieLens-100K and MovieLens-1M (β=1) | MAE: 0.7597 -> 0.7369; 0.7419 -> 0.7275 | Helped |
Surprising findings
- Applying prototype learning directly to NCF (FedProto) gives little improvement, suggesting simple prototype learning is insufficient without item bias modeling.
- PFedRec degrades severely on MovieLens-100K (MAE 1.1708 at beta=1), attributed to over-amplified item learning rate.
Limitations
- Added components increase time and space overhead.
- Lack of semantic interpretation of the item bias embedding and its prototypes.
- Evaluated only on rating prediction with simulated label skew via Dirichlet sampling.
Experimental setup
Datasets: MovieLens-100K MovieLens-1M Amazon-Beauty
Benchmarks: MovieLens-100K MovieLens-1M Amazon-Beauty
Metrics: MAE RMSE
Baselines: FedProx FedMF FedProto FedNCF FedPerGNN PFedRec
Builds on: NCF FedAvg FedProto PFedRec Dirichlet sampling
Closest papers by meaning
- Federated Recommendation with Additive Personalization (ICLR 2024 · similarity 0.86)
- No Prejudice! Fair Federated Graph Neural Networks for Personalized Recommendation (AAAI 2024 · similarity 0.85)
- Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach (AAAI 2025 · similarity 0.84)
- Federated Graph Learning for Cross-Domain Recommendation (NeurIPS 2024 · similarity 0.84)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.83)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · 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.