Novel Machine Learning Techniques for Weather-Related Crash Prediction

This dissertation addresses critical aspects of traffic safety, focusing on novel approaches for weather-related crash prediction—a significant concern in the transportation field. It is divided into three interconnected studies: geospatial risk mapping, the treatment of imbalanced data in machine learning, and analytics for crash prediction. In the first study, the dissertation advances a novel approach to hotspot mapping by developing a spatio-temporal cube that incorporates both the spatial and temporal dimensions of crash data, providing a dynamic and comprehensive analysis of crash hotspots. In the second study, the dissertation tackles the challenge of imbalanced data, which can bias model outputs from the machine learning techniques, making them less adept at predicting crash severity. By extending methods such as Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN), the dissertation evaluates the effectiveness of these methods in datasets with a prevalence of nominal predictors, aiming to enhance the predictive accuracy of machine learning techniques for crash severity. Lastly, the dissertation proposes the use of a Long Short-Term Memory (LSTM) algorithm for predicting a weather-related traffic crash. This approach seeks to overcome the limitations of traditional predictive models by leveraging the ability of LSTMs to retain relevant information over extended time frames, despite the stochastic nature of weather and human behavior.