Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Recommendation with Additive Personalization
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 FedRAP, a federated recommendation method that adds a sparse globally shared item embedding to a private per-user item embedding, with a tanh-based curriculum on regularization weights and differential privacy on the shared matrix.
Problem
Federated recommendation suffers from shared item embeddings that ignore user-specific perception of items, and dense item embedding matrices that cause high communication cost.
Approach
Decompose the item embedding into a sparse global matrix C (L1-regularized, aggregated on server) plus a private user-specific matrix D(i), encouraged to differ via a regularizer, and trained with a curriculum that gradually increases the regularization weights from full to additive personalization.
Key results
- On ML-100K, FedRAP reaches HR@10 of 97.09% and NDCG@10 of 87.81%, versus 70.03% and 41.38% for the best federated baseline PFedRec.
- FedRAP achieves the best HR@10 among federated methods on all six datasets, e.g. 53.98% on QB-article vs 46.00% for PFedRec.
- With (epsilon, delta) differential privacy noise on C, FedRAP's HR@10 on ML-100K drops only 3.55% (0.9709 to 0.9364).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use additive personalized item embeddings rather than a shared-only item embedding (FedRAP-C) (ablation) | FedRAP · ML-100K | HR@10 | Helped |
| Retain shared global item information alongside user-specific item embeddings rather than use local-only embeddings (FedRAP-D) (ablation) | FedRAP · ML-100K | HR@10 | Helped |
| Remove the L1 sparsity constraint on the global item embedding (ablation) | FedRAP · ML-100K | NDCG@10 | No clear effect |
| Replace the L1 constraint with a Frobenius-norm constraint (ablation) | FedRAP · ML-100K | NDCG@10 | No clear effect |
| Use fixed regularization weights instead of gradually increasing weights (ablation) | FedRAP · ML-100K validation | HR@10 | Hurt |
| Use a sine schedule for regularization weights (ablation) | FedRAP · ML-100K validation | HR@10 | Hurt |
| Alternate regularization weights between zero and their maximum every 10 iterations (ablation) | FedRAP · ML-100K validation | HR@10 | Hurt |
| Decrease regularization weights from large to small (fraction schedule) (ablation) | FedRAP · ML-100K validation | HR@10 | Hurt |
| Add Gaussian noise for differential privacy (ablation) | FedRAP · ML-100K | HR@10: 0.9709 -> 0.9364 | Hurt |
| Add Gaussian noise for differential privacy (ablation) | FedRAP · ML-100K | NDCG@10: 0.8781 -> 0.8015 | Hurt |
Surprising findings
- Full-strength additive personalization from the start degrades performance; increasing regularization weights over training (tanh curriculum) works better than fixed weights.
- The L1-sparsified global matrix C has fewer large entries than the L2 variant (4.19% vs 8.07% of parameters above 1e-1), though the L1 and L2 variants perform very similarly.
Limitations
- Each client must store the complete item embedding matrix, which can require significant storage when the item count is large.
- Convergence is slower than PFedRec because the model is more complex.
- Performance degrades under differential privacy noise, though less than baselines.
Experimental setup
Datasets: MovieLens-100K MovieLens-1M Amazon-Instant-Video LastFM-2K Ta Feng Grocery QB-article
Benchmarks: MovieLens-100K MovieLens-1M Amazon-Instant-Video LastFM-2K Ta Feng Grocery QB-article
Metrics: HR@10 NDCG@10
Baselines: NCF LightGCN FedMF FedNCF PFedRec CentRAP
Builds on: FedMF NCF PFedRec
Closest papers by meaning
- Federated Context-Aware Personalized Recommendation (AAAI 2026 · similarity 0.91)
- FedRec++: Lossless Federated Recommendation with Explicit Feedback (AAAI 2021 · similarity 0.89)
- Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach (AAAI 2025 · similarity 0.89)
- Federated Recommendation with Explicitly Encoding Item Bias (AAAI 2025 · similarity 0.86)
- PeFLL: Personalized Federated Learning by Learning to Learn (ICLR 2024 · similarity 0.86)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.86)
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.