Research map / Federated learning and privacy / Privacy attacks and secure learning
HiFi-Gas: Hierarchical Federated Learning Incentive Mechanism Enhanced Gas Usage Estimation
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.
Reports the design and deployment of HiFi-Gas, a hierarchical federated learning incentive mechanism (HFL among gas companies plus VFL between companies and heating stations) with contribution-aware rewards for gas usage estimation at ENN Group.
Problem
Gas usage estimation with federated learning requires incentives so that self-interested gas companies and heating stations contribute high-quality data and compute, while ensuring fairness and resisting low-quality contributions.
Approach
A hierarchical FL framework combining horizontal FL among gas companies with vertical FL between each company and its heating stations, plus a multi-dimensional reward function that splits rewards between data quality (correlation-based score) and model contribution (SMAPE-based leave-one-out style contribution), with a cheaper marginal-contribution estimator than Shapley values.
Key results
- In deployment, average gas usage estimation accuracy rose from 81.1% to 93.0% (+14.67%) for Company A and its heating stations, and from 78.6% to 86.7% (+10.31%) for Company B.
- Offline on ETTh1 without noise, HiFi-Gas took 5.76 s versus about 43.5 s for GTG-S/WT-S/WTDP-S, but its cosine distance (2.69e-2) was worse than the Shapley variants (WTDP-S 4.08e-6).
- Average daily data quantity was reduced by 17.61% under HiFi-Gas, which the authors attribute to participants cleaning their data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Deploy HiFi-Gas for gas usage estimation | Company A and affiliated heating stations · Company A and affiliated heating stations | gas usage estimation accuracy: 81.1% -> 93.0%; +14.67% | Helped |
| Deploy HiFi-Gas for gas usage estimation | Company B and affiliated heating stations · Company B and affiliated heating stations | gas usage estimation accuracy: 78.6% -> 86.7%; +10.31% | Helped |
| Use HiFi-Gas instead of MC for contribution evaluation | ETTh1, different data quantities without noise · ETTh1, different data quantities without noise | cosine distance: 1.31e-1 -> 2.69e-2 | Helped |
| Use HiFi-Gas instead of MC for contribution evaluation | ETTh1, equal data quantity with noise · ETTh1, equal data quantity with noise | cosine distance: 1.70e-1 -> 9.74e-2 | Helped |
| Use HiFi-Gas instead of WTDP-Shapley for contribution evaluation | ETTh1, different data quantities without noise · ETTh1, different data quantities without noise | time (sec): 43.53 -> 5.76 | Helped |
| Use HiFi-Gas instead of WTDP-Shapley for contribution evaluation | ETTh1, equal data quantity with noise · ETTh1, equal data quantity with noise | time (sec): 53.03 -> 5.76 | Helped |
| Use HiFi-Gas instead of WTDP-Shapley for contribution evaluation | ETTh1, different data quantities without noise · ETTh1, different data quantities without noise | cosine distance: 4.08e-6 -> 2.69e-2 | Hurt |
| Use HiFi-Gas instead of WTDP-Shapley for contribution evaluation | ETTh1, equal data quantity with noise · ETTh1, equal data quantity with noise | cosine distance: 9.44e-4 -> 9.74e-2 | Hurt |
| Apply the HiFi-Gas incentive mechanism | ENN Group gas companies and heating stations · ENN Group deployment | average daily data quantity: 17.61% | Hurt |
Surprising findings
- HiFi-Gas was slower-accurate tradeoff: it ran about 7.5x faster than Shapley variants but had higher approximation error than WTDP-Shapley on cosine and Euclidean distance.
- Daily data quantity dropped by 17.61% after deployment, suggesting incentives led participants to prune noisy data.
Limitations
- Offline evaluation uses only 5 clients on the public ETTh1 dataset, not the industrial data.
- Approximation accuracy against the MR Shapley reference is worse than Shapley variants on several metrics.
- Resilience to malicious participants and fairness are left as future work.
Experimental setup
Datasets: ETTh1 ENN gas usage data (historical gas usage, weather, actual usage)
Benchmarks: ETTh1
Metrics: time (sec) cosine distance euclidean distance maximum difference average gas usage estimation accuracy
Baselines: GTG-Shapley WT-Shapley WTDP-Shapley MC
Builds on: FedAvg FedOBD GTG-Shapley WT-Shapley WTDP-Shapley
Closest papers by meaning
- Multi-Session Budget Optimization for Forward Auction-based Federated Learning (ICML 2025 · similarity 0.82)
- Data Disparity and Temporal Unavailability Aware Asynchronous Federated Learning for Predictive Maintenance on Transportation Fleets (AAAI 2024 · similarity 0.81)
- FCOM: A Federated Collaborative Online Monitoring Framework via Representation Learning (AAAI 2025 · similarity 0.81)
- FedRSPO+: A Heterogeneity-aware Algorithm for Decision-focused Federated Learning (NeurIPS 2026 · similarity 0.81)
- Rethinking the Temperature for Federated Heterogeneous Distillation (ICML 2025 · similarity 0.80)
- Towards Diverse Device Heterogeneous Federated Learning via Task Arithmetic Knowledge Integration (NeurIPS 2024 · similarity 0.80)
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.