Comparison of Several Machine Learning Classifiers for Arousal Classification: A Preliminary study

Detection of arousal intervals, especially stress detection via a human-machine interface is a trending topic. Stress detection algorithms with high accuracy can be used in many fields such as criminal interrogations or a variety of stress-related experiments. There are many indicators of the stress on the human body, especially on the face area, such as galvanic skin response (GSR), pupil diameter, heart rate (HR), and electromyography (EMG). Hereby, the measurement of such physiological data in stressful, joyful and non-stressful cases can reveal the effects of the stress on the body signals. This preliminary study aims to compare several machine learning approaches, namely, linear discriminant analysis (LDA), k-nearest neighbour (k-NN), Naive Bayes, support vector machines (SVM) and coarse tree algorithms in a classification study. To perform the analyses, the pupil data are collected from a total of 9 subjects while the subject was watching three types of movies, independently. The classifications are performed among the labelled data with multivariate features such as mean, median, maximum to minimum difference and variance, and their univariate versions in order to observe their independent discrimination performances. Moreover, the preprocessed raw data are also used in classification, independently. Here, the movies are selected such that they include either annotated positive, negative or neutral scenes, which may indicate the stressful, joyful and non-stressful intervals, respectively. Therefore, the classification results of these algorithms for the annotated labels in each channel separately are found to observe their effectiveness in detection of arousal intervals. Hence, the main aim is to contribute to the stress detection literature by providing a comparison between both the classification algorithms, features and raw data classification.