On Machine Learning Applicability to Transaction Time Prediction for Time-Critical C-ITS Applications

Abstract

The demand in providing a reliable operation of modern Cooperative Intelligent Transport Systems (C-ITS) is growing tremendously. One of the enablers for improving efficiency is applying Machine Learning (ML) techniques to predict network characteristics. This work proposes methods to solve the Regression and Classification tasks for the real-life C-ITS mission-critical (transaction) data between buses and the Cloud server collected via three LTE operators. The results show that it is possible to predict the transaction class (classified as (non-)time-critical applications or failed) and the real transaction time with high accuracy. Most applied ML models showed good performance in the Binary Classification of transactions while adding an unsuccessful transactions class (infrequent accuracy). The data imbalance problem arises, resulting in a decrease in balanced accuracy. Moreover, the Gradient Boosting model achieves good results in predicting the real transaction time for a regression task. Numerical results have shown that with the help of ML algorithms with high-quality processing of input data, it is possible to achieve accuracy on a test sample of at least 91.6% in the classification task, and on average, the transaction time error of below 0.08 seconds.