Feature Selection and Model Evaluation Using Machine Learning in Traffic Flow Prediction

Abstract

Urban traffic congestion is a growing challenge, making accurate prediction of traffic flow at intersections essential for optimizing signal control and reducing congestion. This study leverages machine learning techniques to construct a prediction model based on historical traffic flow data. The Pearson coefficient is used to analyze the correlations between numerical features in the dataset, which are then visualized through a heatmap. The Extreme Gradient Boosting (XGBoost) model is employed to train a classifier and evaluate the importance of these features. Ablation experiments are conducted across 14 models—comprising Boosting, Bagging, Linear, and Traditional Lightweight models—using various numerical features. Results show that only three key features, "BusCount," "TruckCount," and "Total," must be retained in the majority of models to maintain stable classification accuracy. Removing these features causes a significant decline in prediction accuracy, with the exception of the Adaptive Boost (AdaBoost) model. This highlights the critical role these features play in effective traffic flow prediction and model stability.