Better understandability of prediction models: a case study for data-based road safety management system

Abstract

The road safety management system aims to ensure a safe transport system for all road users. Analyses of data about traffic accidents can provide important knowledge to support relevant decision-makers or processes. This fact motivated our case study covering the analytical process over publicly available data about traffic accidents in England, Scotland, and Wales. Based on our previous experience with this dataset, we aimed not on the prediction models and their accuracy, but on their explanations for the end-users with limited knowledge from data mining, machine learning, or artificial intelligence. For this purpose, we improved the generated decision models with selected explainable methods like The Local Interpretable Model-Agnostic Explanation (LIME) and SHap Additive exPlanations (SHAP) values. The final visualizations show which attributes and to what extent they contribute to each type of accident.