Road Accident Predictions as a Classification Problem

Abstract

This paper aims at evaluating the performance of various classification methods for road accident prediction. The data is collected under MO- PREVIS [3] project which aims at improving road safety in Portugal. The data is highly imbalanced as there are fewer accident instances than the non-accident ones and due to this imbalance, it is observed that the tra- ditional classification algorithms do not perform well. Using sampling techniques (undersampling and oversampling) improved the results but not significantly. Some methods resulted in increased recall but that de- creased precision as the algorithm returned more false positives to make up for data imbalance.