Road accidents do not generally take place because of a single isolated parameter; in contrast, they occur as a result of the combination of various micro-level driving parameters. The comprehension of the collective influence of such minute parameters assumes paramount importance for the development of a credible predictive model. This paper introduces a machine learning model for the prediction of road accidents employing micro-level driving parameters. The performance of the model is tested on a publicly available database of 2,000 samples of a synthesized driving pattern featuring five driving variables and a target variable defining the accident label [21].
As the dataset has serious class imbalance problems, the Synthetic Minority Over-sampling Technique (SMOTE) is only employed on the training dataset to alleviate bias towards the dominant class [10, 11]. A series of traditional machine learning algorithms, such as Logistic Regression, K-Nearest Neighbor Classifier, Support Vector Machines, and Random Forest Classifier, are trained and compared under the same experimental conditions. The performance of various classifiers is evaluated in terms of accuracy, minority class F1-score, confusion matrix evaluation, and five-fold stratified cross-validation to obtain consistent results [12, 13].
The results of experiments prove that ensemble-based learning with better performance by Random Forest algorithm outperforms baseline models with a remarkable ability to model interaction effects of micro-level driving features [4, 5]. Nevertheless, it is strongly assumed that near perfect results of this study owe little to true generalization performance of an algorithm with its strong reliance on specific patterns encoded within a synthetic environment. As for current results, it must be admitted that these results confirm more of an existence proof rather than an expert judgment of performance of an algorithm for predicting road accidents.