Explainable Stacking-Based Learning Model for Traffic Forecasting

Abstract

This paper implements a two-staged ensemble learning model for traffic forecasting, focusing on the interpretability of predictions. The stacking model leverages the advantages of its diverse component learning models. Experiments on high-dimensional and sparse data validate the stacking model’s superior predictive accuracy compared to baseline models, including LightGBM and XGBoost. In addition to validating the stacking model’s outstanding predictive performance, this paper emphasizes the interpretability of its predictions by proposing an innovative explanation model based on feature contributions. This explanation model addresses the high dimension and sparsity in data prevalent in transportation engineering with its integration of resampling and consensus clustering, offering a scalable, stable, and computationally efficient solution ideal for real-time and large-scale applications. The paper presents theoretical justification, experimental results, and empirical validation of the interpretation model. Extensive experiments demonstrate the model’s enhanced stability compared to traditional shapley additive explanations (SHAP) implementations such as kernel SHAP. Investigating trade-offs between stability and computational efficiency of resampling provides insights for optimal configuration choices. This paper contributes to traffic flow prediction with broad applicability in real-time and large-scale traffic management scenarios, underscoring the vital role of ensemble learning and interpretable machine learning in contemporary data-driven decision making processes.