Research map / Federated learning and privacy / Communication-efficient FL
Ranking-based Client Imitation Selection for Efficient Federated Learning
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 FedRank, a client selection method that casts FL device selection as pairwise ranking, pre-trained via offline imitation learning from analytical selectors (Harmony, Oort, FedMarl) and then refined online with DQN-style RL.
Problem
Client/device selection in federated learning under data and system heterogeneity, where learning-based RL selectors suffer from cold-start and sample inefficiency while heuristics lack generality.
Approach
Combine imitation learning from state-of-the-art analytical selectors for offline warm-start with a pairwise RankNet-style loss so the Q-network learns relative device ordering rather than absolute values, plus an early-exit probing scheme that obtains device states after one local epoch.
Key results
- FedRank improves accuracy by 5.2% (IID) and 11.0% (Non-IID) averaged over OOD datasets, with up to 56.9% gain on the ID MNIST non-IID setting.
- FedRank converges up to 2.01x faster than baselines on ResNet18/CIFAR10 IID training.
- FedRank saves up to 40.1% energy consumption relative to FedAvg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedRank instead of the strongest listed baseline | Client selection for federated learning · MNIST-LeNet5 IID | accuracy (%): 98.84 -> 99(32) | Helped |
| Use FedRank instead of FedMarl | Client selection for federated learning · CIFAR10-ResNet18 IID | accuracy (%): 86.37 -> 87.69 | Helped |
| Use FedRank instead of TiFL | Client selection for federated learning · CINIC10-VGG16 IID | accuracy (%): 67.32 -> 68.58 | Helped |
| Use FedRank instead of Oort | Client selection for federated learning · TinyImageNet-ShuffleNet IID | accuracy (%): 40.57 -> 43.54 | Helped |
| Use FedRank instead of Oort | Client selection for federated learning · MNIST-LeNet5 Non-IID | accuracy (%): 92.47 -> 93.67 | Helped |
| Use FedRank instead of FedMarl | Client selection for federated learning · CIFAR10-ResNet18 Non-IID | accuracy (%): 56.73 -> 59.11 | Helped |
| Use FedRank instead of FedMarl | Client selection for federated learning · CINIC10-VGG16 Non-IID | accuracy (%): 44.16 -> 47.07 | Helped |
| Use FedRank instead of Oort | Client selection for federated learning · TinyImageNet-ShuffleNet Non-IID | accuracy (%): 26.62 -> 30.04 | Helped |
| Add offline imitation learning pre-training (ablation) | FedRank · Non-IID datasets | accuracy | Helped |
| Add pairwise ranking loss (ablation) | FedRank · Non-IID datasets | accuracy | Helped |
| Add probing with early exit after the first local epoch (ablation) | Federated client selection · MNIST-LeNet5; one training round | training latency: 10.6% | Helped |
| Add probing with early exit after the first local epoch (ablation) | Federated client selection · MNIST-LeNet5; one training round | energy consumption: 25.2% | Helped |
Surprising findings
- Training FedRank with RL from scratch (no imitation learning) yields only negligible gains over baselines, attributed to the RL cold-start problem.
Limitations
- Evaluation limited to small image classification models and datasets with simulated heterogeneity plus limited mobile devices.
- Imitation learning pre-training depends on the quality and coverage of chosen analytical experts; generalization to unseen deployments is only shown empirically.
Experimental setup
Datasets: MNIST CIFAR10 CINIC10 TinyImageNet
Benchmarks: MNIST CIFAR10 CINIC10 TinyImageNet
Metrics: test accuracy energy cost per round training speed (time to target accuracy)
Baselines: FedAvg FedProx AFL TiFL Oort Favor FedMarl
Builds on: Oort FedMarl Harmony RankNet VDN
Closest papers by meaning
- Heterogeneity-Guided Client Sampling: Towards Fast and Efficient Non-IID Federated Learning (NeurIPS 2024 · similarity 0.89)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.89)
- FedCompetitors: Harmonious Collaboration in Federated Learning with Competing Participants (AAAI 2024 · similarity 0.88)
- FedImpro: Measuring and Improving Client Update in Federated Learning (ICLR 2024 · similarity 0.87)
- Adaptive Federated Learning with Auto-Tuned Clients (ICLR 2024 · similarity 0.87)
- Efficient Heterogeneity-Aware Federated Active Data Selection (ICML 2025 · similarity 0.87)
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.