**2. Background**

Emotion detection based on physiological data are a vibrant research field that has produced a large body of studies focusing on different emotions and analysis approaches. This section provides an overview of the previous physiology-based emotion detection research relevant to our study. In our literature review, we searched previous studies in Google Scholar with searching keywords of "Emotion", "Boredom", "Classification", "Physiology", and "Sensor". We did not set limitations on the publication year or forum; however, we excluded studies that classified emotions solely by interviews or surveys.

Several AC studies focusing on emotion classification from physiological data [2,6,20–25] referred to Russell's Circumplex model (Figure 1) to explain the target emotion [26]. The model categorizes emotions into four groups by the dimensions of valance and arousal. According to the model, boredom is categorized into the low-valence low-arousal group (red box in Figure 1).

**Figure 1.** Circumplex model [26].

Despite the simplicity of the way in which the Circumplex model assigns boredom to the third quadrant, boredom is considered a complex emotion as various studies have defined it differently. Vogel-Walcutt et al.'s literature review resulted in 37 definitions of boredom, and concluded that "boredom occurs when an individual experiences both the neurological state of low arousal and the psychological state of dissatisfaction, frustration, or disinterest in response to the low arousal." [27]. The conclusions of Russell's model [26] and Vogel-Walcutt et al.'s definition [27] are thus similar. In contrast, the range of boredom in Eastwood et al.'s definition [28] is wider than that of Russell's model. According to their study, people who are in a low-valence state can feel boredom regardless of the level of arousal. Considering these studies, we conclude that a universally accepted definition of boredom does not exist.

Some previous studies regarding boredom categorized it as a trait, while others handled it as a state. The meaning of trait in boredom-related studies is the proneness of an individual to become bored, thus there is a difference between easily becoming bored, and being able to resist boredom. Conversely, the meaning of a state is the current state of boredom that the person is experiencing. Fahlman et al. [29] handled boredom as a trait, while Eastwood et al. [28] and Kim et al. [6] treated it as a state. In this study, we approach boredom as a state. The reason for this is that we hypothesize that, when a person feels bored, changes in their physiological signals can be identified.

Our literature review identified nine studies on boredom classification from physiological data sources as listed in Table 1. It reveals that seven studies used more than one data source; this approach of sensor fusion is a common technique to increase the detection accuracy. The median number of participants in these studies was 21, which is relatively small compared to other cases where machine learning methods are typically applied. In the individual source perspective, EEG was used by three studies, and GSR was utilized by four studies. However, to the best of our knowledge, no previous study has used both EEG and GSR data for classifying boredom.

Sanei and Chambers [30] and Ashwal and Rust [31] showed that EEG data correlates with emotion states of humans. Furthermore, GSR is related to the autonomic nervous system [32], which is also known to correlate with emotion states, thus GSR can be utilized as a potential source for emotion classification [33]. In the physiological perspective, EEG data are captured from the activity of the brain, which belongs to the nervous system together with GSR. Moreover, the analysis on the characteristics of boredom conducted by Bench and Lench [34] suggested that boredom should be associated with the increased autonomic nervous system activity. This linkage implies that a correlation may exist between boredom, EEG and GSR, but so far it has not been investigated in previous studies.

Table 2 presents the methods and the accuracy results of the previous studies that classified boredom using physiological data. Mandryk and Atkins [11], D'Mello et al. [12], Giakoumis et al. [8], and Kim et al. [6] focused on finding correlations between boredom and physiological data using statistical approaches, thus they did not generate classification models. The other reviewed boredom classification studies built classification models using machine learning algorithms and measured the performance of the models. However, these studies did not address the issue of overfitting carefully, making it difficult to guarantee the robustness of the results. Moreover, many previous studies lacked the discussion on the choice of the classification algorithms and only considered a few limited algorithms. Therefore, it is necessary to consider the potential of a wider range of classification algorithms to classify boredom. Considering these facts and shortcomings of previous studies, our study aims to produce reliable performance results based on more diverse machine learning methods.

> **Table 2.** Methods and accuracies of previous boredom classification studies.


#### **3. Data Collection Methodology**

This section describes the methods that we used to collect and analyze physiological data for the classification of boredom. We collected data according to the guidelines of the Declaration of Helsinki [35].

Specifically, we obtained written informed consents from the participants before the data collection, advertised data collection for inviting voluntary participants, explained to the participants that they could quit the experiment anytime they want, and provided snacks as a reward for their participation. The details of the data collection procedure are explained in the following sections.
