Paper Type
Complete
Abstract
Road traffic crashes impose substantial public health and economic costs, yet predicting accident severity remains challenging due to complex, nonlinear interactions among contributing factors. Grounded in sociotechnical systems theory and design science research, this study applies an explainable machine learning framework to 186,101 Florida crash records over seven years, integrating severity prediction, weather analysis, temporal pattern detection, and spatial hotspot identification. We evaluate four ensemble models Random Forest, XGBoost, LightGBM, and CatBoost benchmarked against Multinomial Logistic Regression, employing SHAP to interpret predictions. LightGBM achieved the highest macro F1-score, with SHAP revealing that severity predictions are driven primarily by additive feature effects, particularly visibility and precipitation. Temporal analysis identified commuter-driven patterns peaking during weekday rush hours, while DBSCAN spatial clustering pinpointed concentrated hotspots across Miami, Tampa, and Orlando. Our findings demonstrate that effective traffic safety analytics requires both predictive accuracy and interpretability, offering design principles for intelligent transportation decision support systems.
Paper Number
1896
Recommended Citation
Mahmud, Jishan, "Employing Predictive and Explainable Machine Learning for Traffic Accident Prediction in Florida" (2026). AMCIS 2026 Proceedings. 28.
https://aisel.aisnet.org/amcis2026/sig_dsa/sig_dsa/28
Employing Predictive and Explainable Machine Learning for Traffic Accident Prediction in Florida
Road traffic crashes impose substantial public health and economic costs, yet predicting accident severity remains challenging due to complex, nonlinear interactions among contributing factors. Grounded in sociotechnical systems theory and design science research, this study applies an explainable machine learning framework to 186,101 Florida crash records over seven years, integrating severity prediction, weather analysis, temporal pattern detection, and spatial hotspot identification. We evaluate four ensemble models Random Forest, XGBoost, LightGBM, and CatBoost benchmarked against Multinomial Logistic Regression, employing SHAP to interpret predictions. LightGBM achieved the highest macro F1-score, with SHAP revealing that severity predictions are driven primarily by additive feature effects, particularly visibility and precipitation. Temporal analysis identified commuter-driven patterns peaking during weekday rush hours, while DBSCAN spatial clustering pinpointed concentrated hotspots across Miami, Tampa, and Orlando. Our findings demonstrate that effective traffic safety analytics requires both predictive accuracy and interpretability, offering design principles for intelligent transportation decision support systems.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
SIG DSA