On the Impact of Data Heterogeneity in Federated Learning: A Case Study on Air Quality Prediction

Abstract

Federated learning has become a popular model to apply in privacy-preserving modeling in distributed settings, particularly when the models are applied to data that is distributed among various locations and is often sensitive, such as in air quality prediction. This paper looks into how effective federated learning is for predicting ozone (O₃) concentrations, under both independent and non-independent data distributions. In particular, two representative algorithms Federated Averaging (FedAvg)and Federated Averaging (FedAvg) are experimented on a real-world air quality dataset. An experimental framework was established that was relatively comprehensive and centralized training and local-only models were used as baselines. Mean absolute error (MAE) and root mean squared error (RMSE) are used to measure model performance. The findings suggest that federated learning performs much better than the isolated local models and the performance is similar to that of the centralized training. Having said that, data heterogeneity does present certain issues-it slows down convergence and decreases accuracy in prediction. In such non-IID conditions, FedProx is more stable and less erroneous than FedAvg implying that it is more resistant to client drift.Most importantly, this paper provides empirical data on the impact of data heterogeneity on federated learning, and proves that federated learning can be a viable alternative in privacy-sensitive environmental prediction problems.