Use of information theory measures extracted from OBD-II interface data for driver identification

Abstract

We investigate the use of Machine Learning tools applied to driver identification. We propose us ing Information Theory measures as features in Machine Learning models. The measures used are Statistical Complexity, Permutation Entropy, and Fisher Information. We calculate these mea sures over the raw data of On Board Diagnostics version 2 (OBD-II) sensor values and use them as new features applying to the following classification models: k-Nearest Neighbors, Support Vector Machine, Decision Tree, Random Forest, Multi-Layer Perceptron, and Naive Bayes. We evaluate the driver identification scenario considering sliding windows of 120 samples and an overlap of 60 samples. In each window, the well-established models are trained and evaluated. The models are analyzed by accuracy, by the Area Under the Curve of the Receiver Operating Characteristic Curve (which determines how relevant the classifier is in terms of its sensitivity and specificity), by precision and by recall. We observed significant gains for all models consid ered in the scenario. Following the traditional procedure, we obtained the values: 78.5 to 91.1 % of accuracy, 74.5 to 87.8 % of ROC AUC, 62.5 to 84.2 % of precision, and 60.6 to 85.3 % of recall. While using our proposal, we obtained the values: 94.3 to 99.9 % of accuracy, 91.7 to 99.9 % of ROC AUC, 89.8 to 99.9 % of precision, and 86.4 to 99.9 % of recall.