A robust federated biased learning algorithm for time series forecasting

Mingli Song, Xinyu Zhao, Witold Pedrycz

Introduction

Federated learning (FL) has emerged as a critical paradigm for training machine learning models across decentralized data sources without compromising data privacy. The standard benchmark for this approach, Federated Averaging (FedAvg), aggregates local model updates by simply averaging their parameters. However, this "one-size-fits-all" approach assumes that all local updates are equally reliable. In real-world Internet of Things (IoT) and multi-sensor environments, sensors are often subject to aging, faults, or environmental anomalies, leading to biased or noisy data.

This study introduces Federated Biased Learning (FedBiased), a novel algorithm designed to address the limitations of FedAvg in noisy environments. Beyond privacy, the research explores whether FL offers inherent structural advantages over centralized learning when dealing with data uncertainties.

Methodology

The core innovation of the FedBiased algorithm is a dynamic, performance-based weighting mechanism during the global aggregation step. Unlike FedAvg, which treats every client equally, FedBiased evaluates the reliability of each local model.

The FedBiased Algorithm

The central server in the FedBiased framework calculates a normalized weight for each local model based on its Mean Squared Error (MSE). Models that demonstrate lower error rates are assigned higher weights, ensuring they contribute more significantly to the global model update. This prevents poorly performing or "noisy" local models from degrading the overall system performance.

Flowchart of the FedBiased algorithm illustrating the iterative process of local training and weighted global aggregation.

Model Architecture and Spatial Grouping

For local modeling, the researchers utilized a shallow Back Propagation Neural Network (BPNN) with a sliding window approach (using 24 hours of historical data to predict the 25th hour). While deeper architectures like LSTMs are common for time series, empirical testing showed that the shallow BPNN trained faster and achieved lower MSE for this specific univariate task.

To further optimize the learning process, local clients (monitoring stations) were grouped based on geographic proximity. Euclidean distance calculations involving latitude, longitude, and elevation were used to ensure that the federated groups shared similar environmental contexts.

Experimental Setup

The algorithm was evaluated using hourly Air Quality Index (AQI) data from 35 monitoring stations in Beijing, China, collected during the first quarter of 2022. To test robustness, the researchers created synthetic noisy datasets by injecting Gaussian white noise at various Signal-to-Noise Ratios (SNR = 10, 20, 30, and 40). These tests simulated different scenarios of sensor degradation, including noise across the entire timeline and noise concentrated in specific halves of the training period.

Results and Discussion

The experimental results consistently favored the FedBiased approach over traditional methods.

Superior Accuracy and Statistical Significance

Across various station groupings, FedBiased achieved lower MSE and Mean Absolute Error (MAE) compared to standard FedAvg. Two-sample t-tests confirmed that these performance improvements were statistically significant, with p-values consistently below the 0.05 threshold.

Robustness to Noisy Data

A key finding of the study is the superior robustness of FedBiased in the presence of data corruption. When severe noise (SNR = 40) was injected into a station's data, FedBiased maintained high accuracy (MSE = 0.000136), outperforming both FedAvg (0.000139) and, notably, centralized learning (0.000137). This suggests that federated learning can act as a form of structural regularization, mitigating the impact of localized noise more effectively than a centralized model that pools all data indiscriminately.

Insights on Model Complexity and Scale

The study also highlighted that "more is not always better." Simply increasing the number of stations in a federated group did not guarantee improved performance. Instead, the relevance of the stations (geographic and environmental) proved more critical. Furthermore, the success of the shallow BPNN serves as a reminder that model complexity should be tailored to the specific task rather than defaulting to deep learning architectures.

Conclusion

FedBiased demonstrates that by incorporating accuracy-based weighting, federated learning can become a highly resilient tool for multi-sensor data modeling. The research proves that FL is not just a solution for data privacy but also a robust alternative to centralized learning in real-world, error-prone environments. Future work will focus on integrating spatial relationships more deeply into the model architecture and validating the algorithm across diverse, non-environmental datasets.

Summary updated

Abstract

<title>Abstract</title> The federated averaging algorithm (FedAvg) is extensively used for multi-sensor data modeling but often overlooks the unique characteristics of local models when privacy and data security are not considered. This study introduces a novel federated learning algorithm built upon the FedAvg framework, which emphasizes the specificity of each local model to optimize global knowledge aggregation. The algorithm's effectiveness is demonstrated through an air quality index prediction problem, showcasing superior prediction performance and robustness in noisy data scenarios. Additionally, the study delves into the reliability and robustness of the proposed approach, addressing the prevalent notion that centralized learning methods often surpass federated learning when data security is not a concern. Our experiments affirm the necessity and superiority of federated learning methods, even in the absence of privacy considerations, by effectively managing real-world noisy data.
BibTeX
@article{song2025robust,
  author = {Mingli Song and Xinyu Zhao and Witold Pedrycz},
  title = {{A robust federated biased learning algorithm for time series forecasting}},
  journal = {Clust. Comput.},
  year = {2025},
  doi = {10.1007/s10586-024-05061-7},
}
Figure