A Data-Driven Ensemble Method for Accurate Traffic Data Imputation
Published in IEEE Open Journal of Intelligent Transportation Systems, 2025
Improper handling of missing values in traffic data leads to biased traffic forecasts and poor model performance. Different missing-data conditions require distinct imputation strategies, yet existing methods struggle with complex cases in which multiple conditions coexist. This paper proposes an ensemble learning strategy to adaptively handle different forms of missingness in traffic data imputation. The approach captures intricate temporal and spatial patterns while minimizing imputation errors. Additionally, we introduce a validation pipeline for hyperparameter tuning to optimize model performance. Experiments conducted on two datasets from freeway and urban traffic networks in Belgium show that, for the same base learner, combining models trained under different missingness conditions yields lower imputation error than training under a single condition. The comparison covers twelve base learners, including deep learning, matrix decomposition, and classical methods, and is repeated over multiple mask realizations to confirm that the observed differences are systematic. Results show that commonly used methods such as Simple Imputer and Iterative Imputer from the scikit-learn library consistently yield the poorest performance, highlighting the need for careful method selection in critical applications. While deep learning models achieve strong imputation accuracy, their high computational cost limits their applicability in resource-constrained environments. In contrast, KNN achieves the lowest error under MAR conditions across both regions at negligible computational cost, making it suitable for real-world deployment in edge-based traffic-monitoring systems.
Recommended citation: Morales-Hernández, A., Guastella, D.A., Bontempi, G. (2026). "A Data-Driven Ensemble Method for Accurate Traffic Data Imputation." IEEE Open Journal of Intelligent Transportation Systems.
Download Paper | Download Bibtex
