Evaluation of Multi-Class Multi-Label Machine Learning Methods to Identify the Contributing Factors to the Severity of Animal-Vehicle Collisions

Abstract

Transportation is a fundamental tool to develop communities, cities, and countries on a larger scale, and more extensive transportation networks have developed ubiquitously. However, it is needed to consider the fact that animals also live in the same environment without using the same means, and there is always a chance of colliding with them while driving vehicles. Animal-Vehicle Collision (AVC) is a principal concern for transportation agencies and roadway hazards that influences human safety, property, and wildlife. State of Tennessee animal crash data has been collected for 23 years containing different types of information for each collision. This paper presents and evaluates the performance of five machine learning-based prediction models for animal collisions in the presence of both categorical and non-categorical features. These five models are developed using Logistic Regression, Random Forest, CatBoost, eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LGBM). The CatBoost model has the highest accuracy level at 78.52%. Therefore, it seems to be the most suitable model to predict animal collisions based on 23-year data from Tennessee. The experimental results demonstrate the potential of leveraging categorical data with CatBoost as a viable solution for creating up-to-date and complete analysis for animal-vehicle collision data.

Applying Data Mining Methods to Explore Animal-Vehicle Crashes

“…Some researchers used machine learning models to explore AVC data. Moghaddam et al proposed five machine learning-based prediction models for AVCs in the presence of categorical features, which were developed using eXtreme gradient boosting (XGBoost), logistic regression, CatBoost, random forest, and light gradient boosting machine (LGBM) ( 15 ). The CatBoost model had the highest accuracy (78.52%) and was subsequently the most suitable model for predicting AVCs.…”