Assessing Pedestrian Safety on Roads Through Machine Learning Approaches for State Highways in Washington State

Abstract

The report presents a unique contrast for several Machine Learning approaches aiming at understanding pedestrian fatal collisions. Four classification techniques are applied to assess how roadway features mainly correlate to pedestrian fatal crashes: Logistic Regression, Nearest Neighbor Classification, Decision Tree, and Random Forest Classifier. The data used in this project was collected from the Highway Safety Information System (HSIS) database, which provides both collision data for the entire state of Washington and roadway characteristics for all state highways. Each of the four modeling approaches was implemented using K-fold cross-validation, a process that allows choosing the best parameters for the model. Their results were evaluated and then compared in terms of accuracy score and confusion matrices for the testing data set. It was found that the Decision tree had consistent results and the best performance among all models, showing how the distinct predictors relate to each other to predict fatal pedestrian collisions.