4.2.2. GSR

As mentioned in Section 3.2, we used MV and SR data features from the GSR sensor. Additionally, normalization of MV with feature scaling was also performed (see Equation (2)). As a result, MV, normalized MV (NMV), and SR were secured from the GSR data. Similar to the feature extraction of the EEG data, we calculated the average and the standard deviation for each feature of the GSR data and used them as the final feature values. Consequently, six features in total were secured from the GSR data.

#### *4.3. Machine Learning Model Selection*

Weka, which is an open-source software for data mining that provides several machine learning algorithms, was used for building and testing the machine learning models [40]. In this study, we considered a wide range of machine learning algorithms supported by Weka as candidate algorithms. These candidates were used for initial testing to select the best algorithms for hyperparameter tuning.

Table 3 presents the evaluated algorithms and their options for training. Most of the algorithms were set to default parameters, i.e., the parameters that were preconfigured for the respective algorithms in Weka. Some algorithms (IBk, MLP, SVM) were used several times with different configurations, but these were considered as the same algorithm with different parameters. Therefore, although the number of algorithms was 19, we had 30 models in total to be trained for each dataset (EEG, GSR, and EEG-GSR combined).


**Table 3.** List of algorithms used for training models.

**Network design parameter of MLP (Number of node per layer)**. a = (number of features + number of labels)/2, i = number of features, o = number of labels, t = number of features + number of labels, Ex) if 10 features and 2 labels are used, a, i, o, and t are 6, 10, 2, and 12, respectively (single hidden layer).

IBk, which is a k-Nearest Neighbor classifier, has a parameter for getting a weight from the distance between samples. By default, no weight is assigned. In MLP, the number of layers and the number of nodes in each layer can be defined. For example, "t,i" means that the MLP has two layers, with the first layer consisting of "t" nodes and the second layer consisting of "i" nodes. Table 3 notes explain "a", "i", "o" and "t". Finally, SVM provides options for selecting the SVM kernel type to be used. The MLP, IBk and SVM algorithms also have additional parameters than the ones listed in Table 3; however, we did not adjust them in this study.

#### *4.4. Feature Refinement*

To increase the models' performance, we applied a feature selection algorithm provided by Weka called Wrapper Subset Evaluator (WSE). WSE was proposed by Kohavi and John [41] to find a classifier-optimized feature subset from a dataset to increase model performance. To use WSE, a searching method (forward or backward searching) is required. The target classifier information is also provided as an input to WSE. We used forward searching and the optimized feature set for each algorithm during initial testing and hyperparameter tuning. In the next section, we present the selected features that produced the highest accuracies.
