Abstract
Granular neural networks (GNNs) are a type of prediction models outputting information granules. GNNs not only provide more abstract results and a granular structure but also reveal a flexible nature that can be adjusted by users. As a promising tool, we apply GNNs to solve time series prediction problems under the federated learning (FL) scenario. Distributed time series prediction problems attract more attention recently due to the increased usage of large quantities of IoT sensors and the development of Artificial Intelligence techniques. FL is the main approach to fix the distributed time series prediction problems. In this paper, we design a federated learning framework to refine granular weights of GNNs and then return better prediction results compared with the ones from centralized modeling. Different from studies of FL using numeric neural networks, FL using GNNs is a study of aggregating parameters’ parameters under the federated scenario; thus, the robustness and stability of the method are critical issues. Experiments on predicting the air quality index for 35 stations in Beijing show the effectiveness of our method.
Introduction
In the era of the Internet of Things (IoT), sensors generate massive amounts of distributed time series data. Analyzing this data centrally poses significant privacy risks and high communication costs. Federated Learning (FL) has emerged as a paradigm to train models locally without sharing raw data. However, most FL research focuses on "numeric" neural networks providing point estimates. Real-world data is often noisy; Granular Neural Networks (GNNs) address this by outputting "information granules" (intervals), reflecting human-like reasoning and uncertainty. This paper explores how to aggregate these granular models into a robust global model.
The Federated GNN Framework
The proposed framework integrates GNNs into a federated architecture to refine granular weights across multiple local stations.
The process involves four main stages:
- Local Granulation: Each station starts with a pre-trained numeric neural network. Numeric weights ($W$) are granulated into intervals $[w^-, w^+]$ based on a system-level granularity parameter ($\epsilon$).
- Local Optimization (PSO): Finding the optimal interval for every weight is complex. The framework employs Particle Swarm Optimization (PSO) to find the best set of granularities that maximize a hybrid objective function ($Q$).
- Federated Averaging (FedAvg): Stations transmit their granular weight matrices ($E$) to a central server. The server computes a global granular model ($\bar{E}$) using Federated Averaging and distributes it back to the stations.
- Hybrid Objective Function: The model balances Coverage (percentage of data points within the predicted interval) and Width (precision of the interval).
Experimental Evaluation
The framework was evaluated using the Beijing Air Quality Index (AQI) dataset, comprising data from 35 monitoring stations collected throughout 2022. The task was univariate, one-step-ahead time series prediction.
Key Results
- Performance Gains: The federated approach consistently outperformed centralized modeling. For instance, at a granularity of 0.005, Station 1's objective value improved from 0.0447 to 0.0565 under the FL framework.
- Robustness: Aggregating "parameters of parameters" (uncertainty boundaries) allowed local models to escape local optima, leading to more stable global performance.
- Trade-off Analysis: Using Pareto fronts, the authors identified a "sweet spot" for granularity (e.g., 0.02) where the model maintains high precision while ensuring accurate coverage.
Conclusion
This research demonstrates that federated learning can effectively refine the granular weights of GNNs, providing a robust solution for distributed time series prediction. By sharing model boundaries rather than raw data, the framework preserves privacy while achieving superior predictive performance. Future work will focus on relaxing architectural constraints and improving the interpretability of granular weights.
Summary updated