A Machine Learning Approach for Crossing Accident Prediction: Model Regularization and Optimization

Penalized linear models are commonly used in fatality and disease predictions. Currently, standard approaches for accident prediction are not explored well with zero-inflated sparse traffic data. This manuscript proposed a through research of traffic accident prediction method for penalized linear models with inspiration from elastic net methods. The goal of this study is to extend regularization methods such as ridge, lasso, elastic net in the case of zero-inflated sparse panel data with missing values. Real world accidents-collated data sets are collected from the official website of US Railroad Administration, US Highway Administration and US Police Department from 2000 to 2015. Simulated sampling was applied due to the imbalanced outcome. Models were selected from the Ridge, Lasso, and elastic net model with coordinate descent selections. Multiple solvers were proposed to get the best performance. The final classification model selected by Area Under the Curve (AUC) and f1-score. Comparing to logistics, elastic net improved AUC from 0.41 to 0.48 and f1-score from 0.30 to 0.42 in test data. Comparison with multiple models showed that the elastic net model has high accuracy.