**1. Introduction**

Sensors, machine learning, artificial intelligence, and other kinds of information technologies have recently been advancing rapidly. Based on these trends, several studies have been conducted on the acquisition and processing of information, aiming at a higher-level understanding of the collected information. In particular, detection of the user's context, such as their emotional or physical state, is of particular importance because it enables the creation of context-aware systems that adapt their behavior to match the context in which they are used. This branch of computer science is known as context-aware computing. A sub-branch of it, which is the focus of this study, concentrates on classifying emotions using physiological data.

The study of emotion classification belongs to the area of affective computing (AC) that aims to build computer systems capable of detecting and reacting to the user's emotions. The area of AC in computer science is considered to have been established when a seminal paper by Picard [1] was published, and it has since become a vibrant field of study, with some example studies being [2–4]. To classify emotions in these systems, researchers have three data collection (i.e., measurement) strategies at their disposal [5]: (i) neurological/physiological measurement, which uses sensors to detect changes in the user's body; (ii) subjective self-reporting by questionnaires, diaries or interviews; and (iii) behavioral measurement that is based on expert observations of the participant's behavior. While all these approaches have their specific advantages and disadvantages, as Kim and Fesenmaier [5] suggest, physiological measurement is considered to be particularly objective. In our literature review, we found that a large body of AC studies exists on classifying emotions from physiological data such as electroencephalography (EEG) [2,6,7], galvanic skin response (GSR) [2,8–11], heart rate [2,10,11], and others [12,13]. However, a higher validity can be achieved by combining more than one measurement strategy. For example, a viable approach, which is employed in this study, is to combine a physiological approach with self-reporting, where the latter is used to verify the existence of the target emotion.

Accurate classification of boredom can be considered of particular importance because boredom affects multiple facets of our lives. In a technical report published by the United States Air Force, unmanned aerial vehicle pilots' reaction times were longer when they felt bored [14]. Furthermore, boredom can contribute to serious medical issues such as cardiovascular disease [15]. Additionally, it can have negative effects on learning [16–19]. If computing devices could accurately classify the occurrence of the user's boredom and administer a suitable intervention to compensate for it, they could be used to tackle the aforementioned boredom-related issues.

Several previous studies have built boredom classification models using different physiological data as summarized in Table 1. However, to the best of our knowledge, no previous studies have used both EEG and GSR data for boredom classification. In this study, we performed a joint analysis of both data by collecting EEG and GSR data from 28 participants who also answered a questionnaire surveying their perceived level of boredom. The participants watched two types of video stimuli that were prepared to elicit boredom and to entertain, respectively. Based on the collected data and questionnaire results, we ran an initial test of 19 machine learning algorithms and selected the best three candidate classification models. After hyperparameter tuning, we measured the final performance of the selected models with 1000 iterations of 10-fold cross validation. The best performance with a mean accuracy of 79.98% (min: 71.43%, max: 93.93%) was obtained using a Multilayer perceptron (MLP) model. Furthermore, we analyzed the used features to investigate the correlation between EEG data, GSR data, and boredom. This study, therefore, has three major contributions: (i) revealing a correlation between EEG, GSR, and boredom; (ii) conducting a reliable performance comparison among 19 machine learning algorithms through repeated cross validations; and (iii) proposing a robust boredom classification model based on MLP.


**Table 1.** Studies on boredom classification using physiological data.


HR - Heart Rate, BP - Blood pressure, ECG - Electrocardiogram, PPG - Photo Plethysmo Graphy.
