Comparative Analysis of Machine Learning Models for Predicting Urban Traffic Congestion

Thesis Abstract

Urban traffic congestion remains a persistent challenge for city planners and transportation authorities worldwide, leading to increased travel times, environmental pollution, and economic costs. Traditional traffic prediction models often rely on linear or heuristic approaches that inadequately capture the complex, dynamic nature of urban traffic flow. The advent of machine learning (ML) techniques offers promising alternatives to improve prediction accuracy; however, a comprehensive comparison of their effectiveness in the specific context of urban congestion has yet to be thoroughly conducted. This study aims to evaluate and compare the predictive performance of several prominent machine learning models—including Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting Machines (GBM), and deep learning architectures such as Long Short-Term Memory (LSTM) networks—in forecasting urban traffic congestion. The specific objectives are to determine the models’ accuracy levels using real-world data, analyze their computational efficiency, and identify the most suitable approach for operational deployment in urban traffic management systems. A quantitative research design underpins this study, utilizing observational data collected from a network of traffic sensors installed across a metropolitan city with a population exceeding three million residents. The dataset encompasses traffic volume, speed, and occupancy data collected over a one-year period, comprising approximately 1.5 million recorded instances at five-minute intervals. The sample size was determined through stratified random sampling to ensure comprehensive coverage of different traffic zones and peak/non-peak hours, resulting in an effective sample of 50,000 data points. Data collection instruments include traffic sensors, controller logs, and publicly available traffic reports. Rigorous data pre-processing, including normalization, missing value imputation, and feature engineering, was performed to prepare the dataset for analysis. The study employs multiple analytical techniques aligned with the objective of comparing model performance. Each machine learning algorithm is trained and validated using k-fold cross-validation (k=10) to prevent overfitting and to ensure robustness. Model performance is evaluated through metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²). Additionally, ANOVA tests are conducted to assess the statistical significance of differences observed among model predictions. The theoretical framework grounding this research incorporates the Theory of Complex Systems, which conceptualizes urban traffic as a nonlinear, adaptive system, and the Predictive Modeling Paradigm, which emphasizes the optimization of predictive accuracy through algorithm selection and parameter tuning. Expected findings suggest that deep learning models, particularly LSTM networks, will outperform traditional ML algorithms such as RF, SVM, and GBM in capturing temporal dependencies inherent in traffic data, thereby providing more accurate forecasts. Conversely, models like RF and GBM are anticipated to demonstrate higher computational efficiency, making them suitable for real-time applications where rapid predictions are essential. The comparative analysis is expected to reveal that no single model universally outperforms others across all performance metrics, underscoring the importance of context-driven model selection in urban traffic prediction. This study contributes to the existing body of knowledge by offering a rigorous, empirical comparison of machine learning techniques tailored to urban traffic congestion prediction, highlighting their respective strengths and limitations. It advances the understanding of how different algorithms can be effectively deployed within intelligent transportation systems to mitigate congestion issues. The main conclusion advocates for a hybrid modeling approach that combines the predictive strengths of deep learning with the computational advantages of ensemble techniques. Recommendations for practitioners include adopting LSTM models for high-accuracy forecasting and RF or GBM models for scenarios where speed and resource efficiency are prioritized. Future research should explore the integration of additional data sources, such as weather and social event data, to further enhance prediction accuracy and system robustness.