The Impact of Data Preprocessing on Federated Learning for Air Quality Prediction: A Comparative Study of Imputation Methods

Abstract

Data preprocessing plays a foundational role in machine learning, but it receives limited systematic attention in federated learning (FL) environments. This study empirically compares four missing value strategies using the UCI dataset: direct deletion, mean imputation, K-Nearest Neighbors (KNN) imputation, and Random Forest (RF) imputation. The setup uses a federated framework with five clients, each running a multi layer perceptron model. The findings show direct deletion achieves the strongest performance, with a mean absolute error of 0.173 and an  of 0.966, clearly exceeding imputation methods (the errors around 0.25). The NMHC(GT) feature has an 88.4% missing rate, making imputation unreliable and introducing noise. Although direct deletion reduces the sample from 7,674 to 827 observations, it safeguards data integrity. This research shows that preserving data quality should take priority over quantity when the missing rates are exceptionally high, providing the practical guidance for robust preprocessing design in federated learning systems.