Research map / Federated learning and privacy / Personalized and heterogeneous FL
Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach
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 FedDAE, a federated collaborative filtering method that reformulates CF as a VAE with a global and a local encoder combined by a per-user gating network, and a shared global decoder.
Problem
Federated collaborative filtering under heterogeneous clients: user embeddings alone fail to capture fine-grained personalization, and existing personalized FedCF methods (MF- or item-embedding-based) struggle with non-linear relationships and generalization.
Approach
Decompose user knowledge into a shared global encoder and a private local encoder, mix their Gaussian latent parameters with a data-dependent gating network, and train the whole thing as a VAE with a multinomial-style reconstruction objective that also predicts non-interacted items.
Key results
- On ML-100K, FedDAE reaches HR@20 0.1232 and NDCG@20 0.0503, versus 0.1200 and 0.0501 for the best federated baseline FedVAE.
- On ML-1M, FedDAE achieves HR@20 0.0650 and NDCG@20 0.0261, outperforming all federated baselines (FedVAE 0.0639 HR@20).
- On QB-article and Video, FedDAE is best among federated methods, but performance on these sparse datasets is low and convergence slower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedDAE instead of FedVAE | Federated collaborative filtering · ML-100K | HR@20: 0.1200 ± 0.0040 -> 0.1232 ± 0.0023 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · ML-100K | NDCG@20: 0.0501 ± 0.0032 -> 0.0503 ± 0.0014 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · ML-1M | HR@20: 0.0639 ± 0.0012 -> 0.0650 ± 0.0039 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · ML-1M | NDCG@20: 0.0250 ± 0.0006 -> 0.0261 ± 0.0011 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · Video | HR@20: 0.0634 ± 0.0128 -> 0.0654 ± 0.0080 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · Video | NDCG@20: 0.0227 ± 0.0036 -> 0.0230 ± 0.0029 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · QB-Article | HR@20: 0.0116 ± 0.0031 -> 0.0124 ± 0.0030 | Helped |
| Use FedDAE instead of FedVAE | Federated collaborative filtering · QB-Article | NDCG@20: 0.0045 ± 0.0013 -> 0.0052 ± 0.0014 | Helped |
| Add personalized gating rather than fixed encoder weight w=0.25 (ablation) | FedDAE · ML-100K | HR@20: 0.1213 (5.75%↓) | Helped |
| Add personalized gating rather than fixed encoder weight w=0.5 (ablation) | FedDAE · ML-100K | HR@20: 0.1245 (3.26%↓) | Helped |
| Add personalized gating rather than fixed encoder weight w=0.75 (ablation) | FedDAE · ML-100K | HR@20: 0.1191 (7.46%↓) | Helped |
| Add personalized gating rather than fixed encoder weight w=0.25 (ablation) | FedDAE · ML-100K | NDCG@20: 0.0506 (3.44%↓) | Helped |
Surprising findings
- Fixing the gating weight to 0.75 toward the global encoder degrades performance the most, indicating personalization matters more than shared knowledge.
Limitations
- Performance degrades noticeably on highly sparse datasets (Video, QB-article), with slower convergence.
- Adam optimizer causes fluctuating convergence curves attributed to heterogeneity and noise.
- Privacy protection evaluated only via gradient noise; no formal differential privacy guarantee.
Experimental setup
Datasets: MovieLens-100K MovieLens-1M Amazon-Instant-Video QB-article
Benchmarks: MovieLens-100K MovieLens-1M Amazon-Instant-Video QB-article
Metrics: HR@20 NDCG@20
Baselines: Mult-VAE RecVAE LightGCN FedVAE PFedRec FedRAP CentDAE
Builds on: Mult-VAE beta-VAE FedVAE PFedRec FedRAP
Closest papers by meaning
- Federated Context-Aware Personalized Recommendation (AAAI 2026 · similarity 0.90)
- Federated Reconstruction: Partially Local Federated Learning (NeurIPS 2021 · similarity 0.90)
- Federated Recommendation with Additive Personalization (ICLR 2024 · similarity 0.89)
- DR-VAE: Debiased and Representation-enhanced Variational Autoencoder for Collaborative Recommendation (AAAI 2025 · similarity 0.88)
- Multifaceted User Modeling in Recommendation: A Federated Foundation Models Approach (AAAI 2025 · similarity 0.87)
- Recommendations with Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization (ICML 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.