Research map / Federated learning and privacy / Personalized and heterogeneous FL
FCOM: A Federated Collaborative Online Monitoring Framework via Representation 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 FCOM, a federated collaborative online monitoring method that uses low-rank representation learning and an event-triggered UCB-based allocation of limited monitoring resources across decentralized units.
Problem
Allocating limited monitoring resources across a large population of heterogeneous dynamic processes online, under decentralized data privacy constraints and without a single homogeneous global model.
Approach
Models unit reward functions as combinations of K shared representative models (low-rank representation), estimates them with federated alternating least squares and an event-triggered communication rule, and selects monitored units by a UCB score computed from local statistics only.
Key results
- In the simulation with N=100, K=3, FCOM reached cumulative regret 1383.90 ± 6.90 versus 1764.24 ± 68.64 for LinUCB and 1730.04 ± 155.83 for Sync-LinUCB (CLUCB, centralized, 1335.70 ± 22.89).
- Theory gives regret bound O(NK sqrt(T) ln T) versus O(Np sqrt(T) ln T) for Sync-LinUCB, at the cost of communication O(Nkp ln T) versus O(Np ln T).
- FCOM outperformed LinUCB and Sync-LinUCB on the ADNI MMSE monitoring task under both M=33% and M=66% monitoring scenarios.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FCOM instead of LinUCB | Simulation · Simulation, N = 40 | cumulative regret: 1196.57 \pm 49.03 -> 882.59 \pm 38.79 | Helped |
| Use FCOM instead of Sync-LinUCB | Simulation · Simulation, N = 40 | cumulative regret: 1170.19 \pm 296.36 -> 882.59 \pm 38.79 | Helped |
| Use centralized CLUCB instead of federated FCOM | Simulation · Simulation, N = 40 | cumulative regret: 882.59 \pm 38.79 -> 812.72 \pm 8.75 | Helped |
| Use FCOM instead of LinUCB | Simulation · Simulation, N = 80 | cumulative regret: 1605.92 \pm 32.36 -> 1117.53 \pm 10.78 | Helped |
| Use FCOM instead of Sync-LinUCB | Simulation · Simulation, N = 80 | cumulative regret: 1203.05 \pm 15.08 -> 1117.53 \pm 10.78 | Helped |
| Use FCOM instead of CLUCB | Simulation · Simulation, N = 80 | cumulative regret: 1117.53 \pm 10.78 -> 1061.15 \pm 17.29 | Hurt |
| Use FCOM instead of LinUCB | Simulation · Simulation, N = 100 | cumulative regret: 1764.24 \pm 68.64 -> 1383.90 \pm 6.90 | Helped |
| Use FCOM instead of Sync-LinUCB | Simulation · Simulation, N = 100 | cumulative regret: 1730.04 \pm 155.83 -> 1383.90 \pm 6.90 | Helped |
| Use FCOM instead of CLUCB | Simulation · Simulation, N = 100 | cumulative regret: 1383.90 \pm 6.90 -> 1335.70 \pm 22.89 | Hurt |
| Use FCOM instead of LinUCB | Simulation · Simulation, N = 120 | cumulative regret: 2141.11 \pm 79.36 -> 1553.23 \pm 56.99 | Helped |
| Use FCOM instead of Sync-LinUCB | Simulation · Simulation, N = 120 | cumulative regret: 1808.56 \pm 181.11 -> 1553.23 \pm 56.99 | Helped |
| Use FCOM instead of CLUCB | Simulation · Simulation, N = 120 | cumulative regret: 1553.23 \pm 56.99 -> 1378.40 \pm 70.72 | Hurt |
Surprising findings
- FCOM's regret is only slightly higher than the centralized CLUCB, suggesting the federated privacy-preserving setting costs little performance.
Limitations
- Higher communication cost than Sync-LinUCB (roughly a factor of K).
- Assumes linear reward functions and a shared low-rank representation; asynchronous setting not handled.
- Simulation uses sampled/imputed ADNI data (linear interpolation of missing MMSE values) and repeated only 3 to 5 times.
Experimental setup
Datasets: ADNI MMSE longitudinal data simulated reward functions
Benchmarks: simulation study (N=100, K=3) ADNI Alzheimer's disease MMSE monitoring
Metrics: cumulative regret communication cost
Baselines: LinUCB Sync-LinUCB CLUCB
Builds on: LinUCB Sync-LinUCB CLUCB
Closest papers by meaning
- Streaming Federated Learning with Markovian Data (NeurIPS 2025 · similarity 0.85)
- On the Necessity of Collaboration for Online Model Selection with Decentralized Data (NeurIPS 2024 · similarity 0.83)
- Federated Granger Causality Learning For Interdependent Clients With State Space Representation (ICLR 2025 · similarity 0.83)
- Decentralized Dynamic Cooperation of Personalized Models for Federated Continual Learning (NeurIPS 2025 · similarity 0.83)
- Federated Behavioural Planes: Explaining the Evolution of Client Behaviour in Federated Learning (NeurIPS 2024 · similarity 0.83)
- Communication-Efficient Federated Non-Linear Bandit Optimization (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.