Comparative Analysis of Machine Learning Models for Predicting Urban Traffic Congestion

Thesis Abstract

Urban traffic congestion poses a significant challenge to transportation efficiency, environmental sustainability, and urban livability across many metropolitan areas. Accurate prediction of traffic flow and congestion levels is essential for effective traffic management and urban planning. Traditional methods of traffic forecasting often rely on historical averages or linear models, which frequently fail to capture the complex, nonlinear patterns inherent in urban traffic dynamics. This study aims to conduct a comprehensive comparative analysis of various machine learning models to enhance the accuracy and robustness of urban traffic congestion predictions. The specific objectives include evaluating the predictive performance of models such as Random Forest, Support Vector Regression, Artificial Neural Networks, Gradient Boosting Machines, and Long Short-Term Memory (LSTM) networks; identifying the most effective model for different traffic scenarios; and providing insights into factors influencing model performance within urban traffic datasets. The research adopts a quantitative, cross-sectional research design to facilitate systematic comparison across multiple models. The population comprised urban traffic data collected from the City Traffic Management Authority, encompassing a two-year period (January 2021 to December 2022) with a total of 1.2 million recorded observations. A stratified random sampling technique was employed to select a representative subset of 50,000 data points, ensuring the inclusion of diverse traffic conditions, temporal variations, and geographical zones within the city. Data collection instruments included archived traffic sensor data, GPS-based vehicle movement logs, and weather conditions obtained from municipal meteorological stations, forming a rich multivariate dataset. Data pre-processing involved cleaning, normalization, and feature engineering to enhance model input quality. Model performance was evaluated using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). The analysis employed advanced statistical techniques and machine learning methodologies, including hyperparameter tuning via grid search and cross-validation to prevent overfitting. The study utilized Python-based frameworks such as scikit-learn, TensorFlow, and XGBoost for model implementation. Comparative analysis involved conducting repeated experiments under consistent training and testing partitions to ensure robustness. The predictive accuracy of each model was statistically compared using ANOVA and subsequent post-hoc tests to determine significant differences in performance. Furthermore, feature importance analysis was conducted using permutation methods and SHAP values to interpret model decision-making processes. Expected findings indicate that machine learning models like LSTM and Gradient Boosting Machines are likely to outperform traditional algorithms in capturing temporal and nonlinear dependencies typical of urban traffic patterns. It is anticipated that the study will reveal specific conditions—such as peak hours, adverse weather, or high-density zones—where certain models demonstrate superior predictive capabilities. The analysis is expected to identify the trade-offs between model complexity and interpretability, thereby informing practical deployment considerations in real-world traffic management systems. This research contributes to the existing body of knowledge by providing a rigorous comparative evaluation of contemporary machine learning models tailored for urban traffic prediction, filling gaps related to model robustness across diverse traffic conditions. It advances understanding of the applicability and limitations of different algorithms within the context of large-scale urban data. The findings offer actionable insights for transportation authorities aiming to adopt data-driven congestion management strategies, emphasizing the importance of selecting suitable models based on urban contextual factors. The study concludes that certain ensemble and deep learning approaches, specifically Gradient Boosting and LSTM networks, are most effective for urban traffic congestion prediction. It recommends the integration of these models into real-time traffic monitoring systems, coupled with continuous data collection for adaptive learning. Future research directions include exploring hybrid models that combine strengths of multiple algorithms and extending analysis to include multimodal transportation data for comprehensive urban mobility management.