This article has been accepted for publication in IEEE Access.

This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

Discrimination of Customers
Decision-Making in a Like/Dislike
Shopping Activity Based on Genders: A
Neuromarketing Study
ATEFE HASSANI1 , AMIN HEKMATMANESH2 , AND ALI MOTIE NASRABADI.3
1,3
Department of Biomedical Engineering, Shahed University, Tehran, Iran (e-mail: [email protected])
2
Laboratory of Intelligent Machines at LUT University, Yliopistonkatu 34, 53850 Lappeenranta, Finland (e-mail: [email protected])
Corresponding author: Amin Hekmatmanesh (e-mail: [email protected]).

ABSTRACT The present study considers the decision making of customers in a like/dislike task with
respect to the gender of customers. The investigation is performed by recording electroencephalography
(EEG) signal from 20 subjects that stimulated by displaying images of shoes. In the algorithm, the EEG
signals were denoised by using artifact subspace reconstruction and independent component analysis
methods. The Wavelet technique is then applied to attain five EEG frequency bands and, subsequently
linear and nonlinear features were extracted. The extracted features includes linear features, namely the
power spectral density and energy of wavelet; and nonlinear features, namely the fractal dimension,
entropy, and trajectory volume behavior quantifiers. The meaningfulness of the features for identifying
discriminative channels as well as frequency bands is considered by means of Wilcoxon Rank Sum
statistical test. The identifications of Like/Dislike conditions were then facilitated by the Support Vector
Machine, Random Forest (RF), Linear Discriminant Analysis, and K-Nearest Neighbors classifiers. Results
illustrated that higher frequency bands, the combination of theta, alpha, and beta, in Fp1, Fp2, F7, F8,
Cz, and Pz regions was observed for female group. The most distinctive feature and classifier for the
female group was the energy of the wavelet coefficient and RF classifier, respectively, that produced the
highest accuracy rate of 71.51 ± 5.1%. In addition, the most distinctive features for males were sample
and approximate entropy, as well as the Higuchi fractal dimension that with the RF classifier produced an
accuracy rate of 71.33 ± 14.07%. The linear features investigation revealed more involved brain regions
in a like/dislike task than the previous studies. In addition, it is revealed that the Like decision-making
happens earlier than Dislike.

INDEX TERMS Brain Signal, Like/Dislike, Neuromarketing, Random Forest, Support Vector Machine,
Linear Discriminant Analysis, K-Nearest Neighbors, Wilcoxon Rank Sum

1 I. INTRODUCTION 12 known brands happens on an unconscious level of mind [2].

13 Neuromarketing techniques such as electroencephalography
2 Traditionally, marketing methods (newspaper advertise-
14 (EEG) signal processing has been employed and results
3 ments and television commercials) were invested based
15 are attributed to the human unconscious information during
4 on the customer’s spoken information for identifying their
16 shopping activities. The consequence of the neuromarketing
5 interests. In some points, the investors got success but in
17 is providing informative perspectives for designers about
6 many cases investments failed.
18 specific products and let them understand how products
7 Neuromarketing is a topic for marketing research area that 19 impress the customers and how a decision is made to select
8 uses neuroscience-related techniques to study consumers be- 20 a product. Therefore, success of advertising a product would
9 havior. The concept of neuromarketing was first introduced 21 be increased and costs of advertising would be optimized.
10 by psychologist [1] in the 1990s at Harvard University.
11 It is revealed that the decision-making process about un- 22 Neuromarketing investigations performed by using dif-

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

1 ferent tools, for example Functional Magnetic Resonance 57 Linear Discriminant Analysis (LDA). Additionally, authors
2 Imaging (FMRI) is a technique that measures the brain’s 58 considered the customers preferences in an advertisement
3 metabolic activity; EEG amplifier is technique that mea- 59 task with respect to the background colors of products.
4 sures brain’s electric activity; and eye-tracking technique is 60 Statistical analysis based on the PSD showed the same
5 employed for considering customers attention and focus [3]. 61 results as in [9]. In summary, the obtained locations relative
6 In particular, the FMRI technique is an expensive approach 62 to the willing to buying for a product (Like/Dislike tasks)
7 with measurement ability of high spatial resolution (3 mm 63 were Centro-parietal locations, namely Fp1, Cp3 and Cpz.
8 for the moment) and low temporal resolution (1-3 s for the 64 In addition, significant changes were observed in the frontal
9 moment); EEG technique is an affordable approach with 65 electrodes, namely F4 and Ft8.
10 high temporal resolution in the measurements which is an 66 Aldayel et al. [11] considered customers preferences
11 advantage [4]; and eye-tracking is a technique that examine 67 using EEG signals. In the algorithm, the PSD and valence
12 the participants eyes pupil’s location, but the obtained recent 68 features were extracted from the filtered EEG signals and
13 results found it an unreliable method for the neuromarketing 69 then fed into four different classifiers, named deep learning,
14 studies. Therefore, a numerous neuromarketing studies has 70 SVM, Random Forest (RF) and k-Nearest Neighbor (KNN),
15 been performed by using the EEG signal processing. Here, a 71 that the deep learning provided the best accuracy and
16 selection of related studies to our investigation is presented. 72 precision results. Additionally, results showed that the RF
17 One of the initial neuromarketing studies performed by 73 classifier obtained similar results as in the deep learning
18 Vecchiato et al. [5] that focused on finding relative EEG 74 method. The advantage of the deep learning method is the
19 frequency bands in a Like/Dislike commercial video clip 75 higher potential of producing more accurate and precise
20 task. Authors quantified the decision-making procedure (for 76 results for multi-class identifications problems than the RF
21 Like/Dislike selection) by using power spectral density 77 and SVM classifiers. Also, deep learning has capability of
22 (PSD) brain maps for the theta bands as well as alpha 78 handling a large number of data, which has been applied
23 bands. In addition, a correlation analysis applied on the 79 on a wide range of different biosignal processing in health
24 PSDs of the brain maps. Results showed an asymmetrical 80 monitoring and brain computer interface fields [12], [13].
25 increase for the theta and alpha amplitudes in different 81 The disadvantage of the deep learning algorithm is a large
26 hemispheres when subjects had chosen the pleasant and 82 number of input data is required for the training phase.
27 unpleasant advertisements. 83 Additionally, the training phase is very time consuming
28 Afterwards,Khushaba et al. [6] studied subjects prefer- 84 process.
29 ences to select a product by using the PSD analysis of 85 In another study, Meyerding et al. [14] focused on the
30 brain waves. The spectral activity results showed that the 86 neuronal activation of brain during decision-making of label
31 participants preferences affected the frontal (F3 and F4), 87 brands for different products. In the algorithm, the difference
32 parietal (P7 and P8), and occipital (O1 and O2) areas. The 88 of neurons activities in the prefrontal cortex (PFC) were
33 same team in a series of studies, Khushaba et al. [7] em- 89 measured during subjects coping with two labels brands.
34 ployed a different task for visual stimulation to consider the 90 The achievements based on the fNIRS showed that the PFC
35 subjects preferences again. The obtained results based on the 91 activity for individual subjects increased significantly. The
36 PSD features showed significant changes in the following 92 main limitation of the study was participating a low number
37 regions: frontal region (delta, alpha, and beta across F3, 93 of subjects in the experiment.
38 F4, Fc5, and Fc6), temporal region (alpha, beta, gamma 94 Most recent neuromarketing studies focused on the com-
39 across T7), and occipital region (theta, alpha, and beta across 95 bination of subjects, while gender is an important parameter
40 O1). Later on, Korkmaz et al. [8] used frequency contents 96 which has not been taken into account in the investigations.
41 for determination of the most discriminative channels and 97 The primary functions of human’s perception is various
42 frequencies in a Like/Dislike neuromarketing study. Results 98 for different genders points of view and it would be an
43 showed that the frequency bands of 4 Hz and 5 Hz were 99 impressive parameter for improving the neuromarketing
44 determined as the most discriminative ranges in the left 100 models, which has not been considered yet. In addition,
45 frontal (F7) and right temporal (T6) regions. 101 linear features are exhaustively studied in the computational
46 In a different study, Ramsoy et al. [9] performed a neu- 102 parts and a lack of nonlinear feature investigation exists.
47 romarketing investigation of the mechanisms of customers 103 Table. 1 present a review of some experimental neuromar-
48 decision-making for willing to pay for a product. In the 104 keting studies.
49 procedure, the alpha, beta, and gamma frequencies of EEG 105 The first contribution of the present study is identifying
50 signal were considered. Statistical analysis showed that 106 the most discriminative areas of brain in a like/dislike
51 the prefrontal gamma asymmetry was related to willing to 107 (willing to buy) task with respect to the subject’s gender as
52 pay responses significantly. Then, Golnar-Nik et al. [10] 108 an important factor. The second contribution is considering
53 considered the hypothesis if the PSD feature from the EEG 109 the EEG frequency bands in the determined regions and
54 is a suitable approach for predicting the customers decision- 110 investigating linear and nonlinear features for identifying
55 making in a like/dislike task. In the algorithm the features 111 Like/Dislike conditions by using the RF, SVM, LDA, and
56 were classified by using Support Vector Machine (SVM) and 112 KNN classifiers with respect to the gender of participants.
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

TABLE 1: EEG-based neuromarketing studies for assesing the customer preference.

Year Reference Number of Subjects Class Feature Extraction Selection Classification Algorithm
1. Like
2013 [8] 15 (10 females and 5 males) Power spectral density -
2. Dislike
Time-frequency analysis
1. Like Frequency domain:
2016 [15] 5 (3 females and 2 males) SVM, KNN
2. Dislike Power spectral density,
Band power, Spectrum power
1. Like ANN (2 layers + Sigmoid),
2017 [16] 40 (25 males and 15 females) Wavelet analysis
2. Dislike SVM, RF, HMM
SVM, KNN,
1. Most favored
2018 [17] 33 (13 males and 20 female) Power Logistic Regression,
2. Least favored
Decision trees
1. Like
2019 [10] 16 (9 males and 7 females) 2. Dislike Power SVM, LDA
3. Buy
1. Like
2020 [11] 32 Power spectral density, Valence DNN, KNN, SVM, RF
2. Dislike
1. Like
2021 [18] 25 Spectral energy Ensemble Classifier, DNN
2. Dislike
1. Positive affective attitude
2022 [19] 20 Time, Frequency, and Time-frequency SVM
2. Negative affective attitude

1 The Like/Dislike decision-making interpreting as a willing 37 images in the MATLAB software simulation environment.
2 to pay for a product or not. The employed features included 38 The structure of experimental protocol is shown in Fig. 1.
3 as follows: PSD and energy of wavelet coefficients, fractal
4 dimension, entropy, and trajectory volume behavior quan-
5 tifiers. The rest of paper is presented as follows: Section Stimuli (15 s)

6 II is the experimental setup and data acquisition; Section

7 III is mathematical methods; Section IV is the results and Rest (10 s)

8 discussion; and section V is the conclusion.

Stimuli (15 s)
9 II. EXPERIMENTAL SETUP
10 The EEG signals recorded from twenty (10 males and 10
11 females) healthy and graduated right-handed students. Be-
12 fore the experiment, the required regulations and conditions
13 for the task were explained to the subjects and a consent
14 of participation was signed. Participants had no history of
15 neurological or psychiatric disorders as well as no record
16 of medication, alcohol, and drugs consumption. Subjects FIGURE 1: The employed structure of experimental protocol in
the experiment.
17 were recommended not to use caffeinated and nicotine
18 substances for at least four hours before the experiment.
19 The experimental procedure was approved by the national 39 A. DATA DESCRIPTION
20 ethics committee in biomedical research. 40 A 16-channel EEG amplifier (g.USBamp, g.tec, Austria)
21 During the EEG signal recording, subjects should sat on a 41 was used to record electrical potential. The EEG electrodes
22 comfortable chair in a dimly lit room. A set of stimulus im- 42 were installed based on the international 10-20 electrode
23 ages were displayed for the subjects. The images consisted 43 location system, in which enable us to cover almost all
24 of 16 women’s shoes with different styles and colors. Each 44 areas of a head. The main purpose of the 10-20 standard
25 image displayed 15 seconds on a screen which is located 45 is providing electrode instalment using a small number of
26 at a distance of one meter. To choose one pairs of shoes, 46 electrodes (typically 21 electrodes) for recording EEG. In
27 an assigned key should be pressed by the subjects for like 47 our experiment, the right ear and Fpz electrodes were set
28 (willing to pay) the shoes or another assigned key should 48 as the GROUND (GND) and common reference for all the
29 be pressed for dislike the shoes (not willing to pay). In the 49 channels, respectively. Our electrode instalment is showed in
30 task, a ten second interval rest between displaying images 50 Fig. 2. The EEG was recorded with the sampling frequency
31 (shoes) was set. The time responses were recorded for the 51 of 256 Hz, and a High-pass filter with a cut off frequency of
32 further process. The displayed pictures were women shoes 52 0.1 Hz and a notch filter with a cut off frequency of 50 Hz
33 and the main question for women was are you willing to pay 53 were applied. Table. 2 shows the number of decisions for
34 for the shoes for her self and male subjects should answer 54 subjects. After collecting the EEG data, the preprocessing
35 to the question are you willing to pay for the shoes for your 55 method is applied to remove noise.
36 life partner (wife). Paradigm blocks were used to display
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

27 2) Independent Component Analysis

28 In the second noise removal step, an ICA algorithm applied
29 on the denoised signal. The basic hypothesis concept of ICA
30 is the EEG signal is weighted by using linear combination of
31 electrical potentials, which are generated from independent
32 brain sources [22], [23]. The goal of ICA algorithm is
33 finding the least Gaussian state in a new space in which
34 leads to the identification of the main sources. The ICA
35 is not able to separate the sources of noises if they are
36 completely Gaussian. If the computed components were not
37 independent, then the ICA seek for the most independent
38 space to separate linear sources [24]. Therefore, we mapped
39 the EEG data from the sensor space into the source space
40 and removing the major interference sources including eye
41 components, muscle components, heart components, as well
FIGURE 2: Representation of 10-20 system for electrode place- 42 as line noise, and channel noise [25].
ment.
43 B. FEATURE EXTRACTION

TABLE 2: Number of "Like" and "Dislike" decisions for 44 To identify the active brain regions in our marketing stimu-
male and female subjects. 45 lus experiments, different linear and nonlinear features from
46 five different frequency bands were extracted and analyzed.
Number of decisions 47 These analysis were applied on three groups with respect
Like Dislike 48 to the genders (male, female and a combination of male
1 49 and female). To this end, two sets of linear and nonlinear
Male 71 89
50 features were extracted from the EEG signals.
Female 83 77
Total 154 166
51 1) Linear Feature Extraction

2 III. MATHEMATICAL METHODS 52 The linear feature extraction approaches includes the energy
3 Here, the employed mathematical methods for analysing the 53 of wavelet coefficients and PSD. To this end, the EEG
4 EEG signals are presented.The procedure of data analysis 54 signals were decomposed by using the Discrete Wavelet
5 is shown in Fig. 3. The first step to process the EEG signal 55 Transform (DWT) algorithm to obtain different frequency
6 is preprocessing. 56 bands. Therefore, a mother wavelet is set in the DWT,
57 named Daubechies 8 (db8) and then the DWT was applied
58 for five levels to reach the aim frequency bands [26], [27].
7 A. PREPROCESSING
59 The energy of DWT coefficients were computed as follows:
8 EEG signals are susceptible to be contaminated easily
9 with physiological and environmental artifacts. To remove N
10 noises, the related EEG segments to the decision-making Ei =
X
Ci , j 2 , (1)
11 events (Like/Dislike) were extracted. Then, different de- i=1
12 noising algorithms were applied to the segments, namely
13 Artifact Subspace Reconstruction (ASR) and Independent 60 where Ci , j represents i-th coefficient of j-band and N
14 Component Analysis (ICA). The main advantage of above- 61 is the number of j-band coefficients.
15 mentioned denoising methods is well-handling physiological 62 The PSD feature computed by means of modified peri-
16 artifacts with less signal alterations in comparison with 63 odogram algorithm, which is a non-parametric estimation.
17 the frequency-based filters. The concept of ASR and ICA 64 The PSD algorithm involves following steps: 1) multiplying
18 methods are presented as follows: 65 the input time series with a non-rectangular window such
66 as Hamming; 2) applying the Fourier transform function;
19 1) Artifact Subspace Reconstruction 67 and 3) computing the density of power spectrum estimation
20 The first noise removal step was employing an ASR method. 68 by using the Fourier transform size. Typically, the non-
21 ASR is one of the automatic artifact removal methods, which 69 parametric methods have less computational complexity
22 is proposed by Kothe et al. [20]. The ASR is an effective 70 than the parametric models. Therefore, the PSD features
23 component-based algorithm for removing transient artifacts 71 were extracted from the frequency bands consist of Delta
24 in a real-time system [21]. In our algorithm, to remove noise 72 (0.5-4 Hz), Theta (4-8 Hz), Alpha (8-13 Hz), Beta (13-30
25 components we employed the well-known ICA method after 73 Hz), and Gamma (30-40 Hz). In the computations, there was
26 applying the noise removal ASR technique. 74 no restriction for selecting the frequency bands.
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

Focus & Rest Focus &

Decision Decision

Filtering
Bandpass (0.1-50 Hz)

preprocessing
ASR, ICA

Extracting frequency
bands

Extracting Features
Statistical Analysis
Linear & Nonlinear
Wilcoxon signed rank test
features

Classification
RF, SVM, LDA, KNN

FIGURE 3: Procedure of data analysis of the proposed system.

1 2) Nonlinear Feature Extraction 35 According to the Taken’s theory [33], growing the tra-
2 Brain function is a non-stationary system that produces 36 jectory of EEG signals in a reconstructed phase space is
3 signals with high level of complexity [28]. Since now several 37 applicable using the time series X(t) = {x1 , x2 , ..., xN }
4 methods has been developed to interpret the brain neurons 38 and two delayed input signals as follows [34], [35]:
5 activities which are recorded by the EEG. One approach to
6 study the dynamics of a nonlinear system is investigating →
−
X i (t) = (xi , xi+τ , ..., xi+(µ−1)τ ) i = [1 N −(µ−1)τ ]
7 the chaotic behavior of a system [29]. The chaotic methods
(2)
8 apply techniques for quantifying the nonlinear features of a
39 where µ and τ are embedding dimension and delay,
9 system such as quantification of trajectory volume behavior,
40 respectively, which are obtained based on the false nearest-
10 entropy and fractal dimensions [26].
41 neighbors and mutual information methods. To extract fea-
11 The idea of entropy was first proposed in the thermody- 42 tures, the Euclidean distance matrices (T ) of trajectories (L)
12 namic field to measures the trajectory of a system. The en- 43 with length N between all the state vectors are computed
13 tropy concept describes the behavior of a part of a trajectory 44 follows:
14 that can be predicted from the rest. In the computations, the
15 higher entropy in a system leads to higher complexity, which →
− → − →
−
L = X 1 , X 2 , ..., X N . (3)
16 means less chance for predicting the systems behaviour
′
17 [30], [31]. In the present study, we have employed different 45 Matrix T is obtained by removing the main diagonal of
′
18 features based on the entropy computations, namely spectral 46 matrix T . The last column of the matrix T is then removed
′′
19 entropy, approximation entropy, and sample entropy. 47 and shifted to the left. Matrix T with dimension (M ×(M −
20 The second set of our chaotic features for measuring 48 2)) represents the difference between the distances of state
21 the complexity of an EEG was fractal dimensions. The 49 vectors and trajectory motion.
22 fractal dimension concept is a numerical scale method for 50 Regarding the trajectory of EEG signals in the recon-
23 measuring how much space is filled by a pattern. The 51 structed phase space several features extracted, namely
24 hallmark characteristics of fractals that we make use of 52 Occupational Size (OS), Average Expansion Speed (AES),
25 them are self-similarity and non-integer dimension. In our 53 Average Compression Speed (ACS), Average Expansion
26 computations, Higuchi and Katz fractal dimensions were 54 (AE), Average Compression (AC), Standard Deviation Ex-
27 employed [31]. 55 pansion Speed (SDES), Standard Deviation Compression
28 The third set of chaotic features is the trajectory of EEG 56 Speed (SDCS), and Complexity. The above-mentioned fea-
29 signals in a reconstructed phase space, which is computed 57 ture definitions are as follows:
30 using two delayed EEG signals [29]. The obtained chaotic- 58 • OS: The average across all elements in matrix T .
′′
31 based quantifiers quantify the volumetric behavior of a 59 • AES: The average of negative numbers in matrix T .
′′
32 trajectory in a phase space. The basis of computations is the 60 • ACS: The average of positive numbers in matrix T .
33 expansion and compaction of a trajectory. For this purpose, 61 • AE: The normalized expansion rate.
34 the features in [32] study are implemented. 62 • AC: The normalized compression rate.

VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

′′
1 • SDES: Variation of negative numbers in matrix T . 53 reducing the probability of overfitting and high variance
′′
2 • SDCS: Variation of positive numbers in matrix T . 54 in comparison with decision tree algorithms. The second
3 • Complexity: Summation of number of the positive 55 classifier is SVM method.
′′
4 elements in SDCS (T ) and the number of negative
′′
5 elements in SDES (T ) after normalization. 56 2) SVM Classifier
6 The computed features are analyzed to find the most 57 SVM is a supervised learning classifier, which is based on
7 discriminative channels and frequency bands between the 58 statistical learning theory. The first nonlinear SVM model
8 Like/Dislike groups. To evaluate the separability of features, 59 introduced by Vapnik [38]. Conceptually, the SVM model
9 the Wilcoxon Rank Sum statistical test is applied, which is 60 is a hyperplane or line that separates a set of positive and
10 explained in the next step. 61 negative samples, namely support vectors, by using a max-
62 imum distance [39], [40]. Recently, several modifications
11 C. STATISTICAL ANALYSIS 63 of the SVM has been developed to increase the accuracy
64 and precision results [41], [42]. In the computations, the
12 The Wilcoxon Rank Sum statistical analysis is a non-
65 boundaries of the two classes may not be linearly separable.
13 parametric test that examines the difference of a mea-
66 Therefore, features are mapped from the input space to a
14 surement quantity between two groups whose samples are
67 feature space with higher dimension by means of nonlinear
15 independent from each other. This test is based on the
68 kernels. The mapped feature space dimension is increased
16 median comparison. In this research, the Wilcoxon Rank
69 until the features are separable linearly [43]. Different types
17 Sum statistical analysis was applied for recognition of the
70 of kernels has been employed for the SVM decision func-
18 discriminative channels in five frequency bands. The ob-
71 tion, the most common used kernel in the SVM algorithm
19 tained features distribution was not Gaussian and it satisfies
72 is the Radial Basis Function (RBF). The advantages of RBF
20 the Wilcoxon Rank Sum’s assumptions. We used the MAT-
73 are using Gaussian shape for the distributions and enabling a
21 LAB statistical analysis toolbox for the computations. In our
74 feature space with unlimited dimensions [44]. Also, several
22 experiment, the statistical analysis between the difference
75 method developed to improve the capability of RBF [41],
23 of Like/Dislike groups found significant. The determined
76 [45].
24 significant features are then employed for identifying the
25 Like/Dislike conditions by using RF, SVM, LDA, and
77 3) LDA Classifier
26 KNN classifiers. SVM and LDA are the most common
78 LDA is a linear model commonly used for supervised
27 metrics used by most research in the domain of customer
79 classification problems and dimensionality reduction. This
28 preferences for evaluation. More than half of the EEG-
80 technique offer a good separation between different classes
29 based papers have used them [11]. According to reference
81 and avoid overfitting. As a result, computational costs will
30 [11], deep learning has similar results to RF. So, we used
82 be significantly reduced and classification will be more
31 Random Forest rather than DNN due to the limited number
83 accurate by projecting the given n-dimensional feature space
32 of subjects.
84 onto a smaller feature space while maximizing the class
85 separability [46].
33 D. CLASSIFICATION
34 To classify the Like/Dislike conditions several studies have 86 4) KNN Classifier
35 been employed by using linear features. In addition, we have 87 KNN is a non-parametric and supervised learning algorithm.
36 extracted discriminative frequency bands by using classifi- 88 This algorithm is simple to implement and robust to the
37 cation. Here, we employed the well-known RF, SVM, LDA, 89 noisy training data which can be used for both classification
38 and KNN classifiers to identify the Like/Dislike conditions 90 and regression. In KNN, inputs are classified based on their
39 by using linear and nonlinear features. 91 K neighbors. The disadvantage of the algorithm is the value
92 of K will always need to be determined, which may be
40 1) RF Classifier 93 complicated. Since the distance between the data points for
41 RF is a statistical method for a supervised learning classifier, 94 all the samples of training dataset must be calculated, the
42 which is introduce by Leo Breiman [36]. The RF makes 95 computation cost will be high. In the next step, the obtained
43 several decision trees and merges them together to make 96 results are presented and discussed.
44 a more accurate and stable prediction. The RF classifier
45 generates trees randomly to generate forests. The greater the 97 IV. RESULT AND DISCUSSION
46 number of trees, the more accurate estimation. Therefore, 98 In the present study, 20 subjects participated in the ex-
47 it is concluded that a meaningful relation between the 99 periment and EEG signals were recorded by using visual
48 number of trees in the RF algorithm and the obtained results 100 Like/Dislike stimulation task. The above-mentioned algo-
49 exists [37]. Shortly, the RF contains multiple decision trees 101 rithms were then applied on the EEG data to find the
50 that each decision tree considers a set of data. The final 102 distinctive frequency bands and the relative areas of brain
51 RF classification decision is made based on the previous 103 which were affected by willing to pay decision for a pair of
52 decisions in the trees. The advantages of RF algorithm is 104 shoes. Here, the results of frequency bands, brain regions
6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

1 and the effects of nonlinear and linear features on the 56 (average Like > Dislike): the delta band in right frontal
2 distinguished brain regions are analyzed with respect to the 57 (F8), left parietal (P3) and central parietal (Pz) regions; and
3 gender of subjects. 58 the alpha band in left frontal (F3). In addition, the activated
59 frequency bands and regions underlying of Dislike condition
4 A. STATISTICAL ANALYSIS OF LINEAR FEATURES 60 (not willing to pay) are as follows (average Dislike > like):
5 VS. NONLINEAR FEATURES 61 the theta band in right frontal (F4) region; the alpha band in
6 To investigate how gender influences decision-making, 62 right parietal (P4) region; and the beta band in right frontal
7 subjects were categorized in three groups to perform a 63 (F8) region. Moreover, results showed that the extracted
8 Like/Dislike task. The groups involved 1) males, 2) females 64 linear features from the delta, theta, alpha, beta, and gamma
9 and 3) combination of males and females. In addition 65 frequency bands for identifying the Like/Dislike conditions
10 to using linear features, a set of nonlinear features were 66 were insignificant. As it is evident, the nonlinear features
11 employed to compute features. The Wilcoxon Rank Sum 67 has potential of carrying information from the EEG which
12 statistical analysis was then applied on the explained linear 68 are not visible by using the linear attribute features.
13 and nonlinear features, which were extracted from the main 69 3) Combination of male and female groups: Fig. 6a
14 five frequency bands and then scalp maps were created. The 70 demonstrates the discriminative channels and frequency
15 guide line for interpreting scalp maps such as Fig. 4b is as 71 bands based on nonlinear features between Like/Dislike
16 follows: the blue regions shows that the average value of 72 conditions. It is evident that the activated frequency bands
17 extracted features for the Dislike condition is greater than 73 and regions underlying of Like condition (willing to pay) are
18 the Like condition. Additionally, the red regions are the 74 as follows (average Like > Dislike): the delta band in right
19 channels where the average value of extracted features for 75 frontal (F8) region; the beta band in left prefrontal (Fp1)
20 the Like condition is greater than the Dislike condition. The 76 and left occipital (O1) regions. In addition, the activated
21 following parts are analyzing the obtained results in three 77 frequency bands and regions underlying of Dislike condition
22 parts (significant features, channels and frequency ranges) 78 (not willing to pay) are as follows (average Dislike > like):
23 for each group. 79 the theta band in right frontal (F4) region; and the gamma
24 1) Consideration of female group: Fig. 4a demonstrates 80 band in left central (C3) region. For all subjects, the alpha
25 the discriminative channels and frequency bands based on 81 band changes between Like/Dislike conditions was insignif-
26 nonlinear features between Like/Dislike conditions. It is evi- 82 icant.
27 dent that the activated frequency bands and regions underly- 83 Fig. 6b demonstrates distinctive regions in all frequency
28 ing of Like condition (willing to pay) are as follows (average 84 bands based on the linear features between Like/Dislike
29 Like > Dislike): the theta band in parietal (Pz) region; the 85 decision-making task. It is evident that the average of
30 alpha band in right prefrontal (Fp2) region; the beta band in 86 Dislike condition was stronger than the Like condition. The
31 right frontal (F8) region; and the gamma band in left central 87 significant results for Dislike condition were obtained from
32 (C3) region. Additionally, the activated frequency bands and 88 the theta band in left central (C3) region. The obtained
33 regions underlying of Dislike condition (not willing to pay) 89 results based on the nonlinear features showed insignificant
34 are as follows (average Dislike > Like): the theta band 90 differences in the delta, alpha, beta, and gamma frequency
35 in left frontal (F3) region; the beta band in left occipital 91 bands between Like/Dislike conditions.
36 (O1) region; and the gamma band in central (Cz) region.
37 Fig. 4a shows non-distinctive changes by using the nonlinear 92 B. LIKE/DISLIKE CLASSIFICATION PERFORMANCE
38 features from the delta band between Like/Dislike classes. 93 To identify the discriminated frequency bands between
39 Fig. 4b demonstrates the discriminative channels and 94 Like/Dislike groups, well-established RF, SVM, LDA, and
40 frequency bands based on the linear features between 95 KNN classifiers were employed. In the process, a matrix size
41 Like/Dislike conditions. It is evident that the average of 96 of 15 × 14 × 160 was formed for individual subjects, i.e.,
42 Dislike condition was stronger than the Like condition (blue 97 15 channels, 14 features, and 160 Like/Dislike epochs. In
43 regions - average Dislike > Like). The affected frequency 98 our study, the reported accuracy results for the Like/Dislike
44 bands and regions for the Dislike condition are as follows: 99 classification are based on a 10-fold cross-validation. The
45 the theta band in right central (C4) region; the alpha band 100 explanations are performed based on the genders.
46 in frontal (F7, F8, Fp2) and right central (C4) regions; the 101 Performance of female group classification: table. 3a
47 beta band in left frontal (Fp1, F7) and parietal (Pz) regions. 102 illustrates the performance of the Like/Dislike conditions
48 No region was affected under Like condition by using the 103 identification based on the linear features. According to
49 linear features. Results showed insignificant changes in the 104 the RF, SVM, LDA, and KNN classifiers results, the SVM
50 delta and gamma bands. 105 classifier achieved the highest accuracy of 71.51 ± 5.10% .
51 2) Consideration of male group: Fig. 5 demonstrates 106 In the algorithm, the most informative revealed extracted
52 the discriminative channels and frequency bands based on 107 feature was the energy of wavelet coefficient from the
53 nonlinear features between Like/Dislike conditions. It is 108 combination of theta, alpha, and beta frequency bands in
54 evident that the activated frequency bands and regions 109 channels Fp1, Fp2, F7, F8, Cz, and Pz. Table. 3b illustrates
55 underlying of Like condition (willing to pay) are as follows 110 the results of Like/Dislike conditions identification based on
VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

(a) Nonlinear features for female subjects

(b) Linear features for female subjects

FIGURE 4: The obtained affected areas of female’s brain during displaying images in a Like/Dislike task, by using (a)
Nonlinear and (b) linear features.

FIGURE 5: The obtained affected areas of male’s brain during displaying images in a Like/Dislike task by using nonlinear features.

(a) Nonlinear features for the combination of all subjects.

(b) Linear features for the combination of all subjects.

FIGURE 6: The obtained affected areas of all subject’s brain during displaying images for a Like/Dislike task, by using (a)
nonlinear and (b) linear features.

8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

TABLE 3: Like/Dislike condition classification by using (a) linear and (b) Nonlinear features for female subjects.
(a) Linear features for female subjects.
Classifier Metric Theta Alpha Beta All Bands
Accuracy 55.11 ± 7.19 61.91 ± 5.44 58.31 ± 8.81 63.19 ± 7.02
RF Sensitivity 57.76 ± 15.63 65.77 ± 11.53 60.92 ± 9.22 64.21 ± 7.46
Specificity 53.79 ± 7.71 59.84 ± 11.71 56.55 ± 10.98 62.66 ± 12.1
Accuracy 65.53 ± 5.84 57.94 ± 3.34 60.81 ± 2.39 71.51 ± 5.10
SVM Sensitivity 57.18 ± 6.85 55.02 ± 4.51 63.85 ± 7.41 78.19 ± 5.79
Specificity 74.27 ± 6.72 60.79 ± 6.76 57.75 ± 7.92 64.88 ± 6.16
Accuracy 61.08 ± 3.73 56.75 ± 0.02 62.44 ± 3.72 62.14 ± 2.17
LDA Sensitivity 54.45 ± 4.79 51.42 ± 5.69 72.86 ± 6.08 56.20 ± 7.01
Specificity 67.72 ± 3.88 61.94 ± 2.87 52.17 ± 4.29 67.8 ± 3.94
Accuracy 60.36 ± 8.73 61.96 ± 3.24 60.03 ± 4.24 61.84 ± 3.95
KNN Sensitivity 55.23 ± 7.45 63.2 ± 2.94 57.29 ± 3.51 69.4 ± 1.21
Specificity 65.44 ± 10.72 60.71 ± 5.46 62.75 ± 5.78 54.7 ± .58

(b) Non-Linear features for female subjects.

Classifier Metric Theta Alpha Beta Gamma All Bands
Accuracy 55.53 ± 9.59 57.87 ± 6.41 55.74 ± 10.31 61.87 ± 10.39 60 ± 8.32
RF Sensitivity 53.73 ± 16.19 60.19 ± 8.82 52.66 ± 13.27 65.81 ± 10.24 54.28 ± 10.37
Specificity 57.64 ± 11.66 55.97 ± 8.44 59.94 ± 11.27 59.25 ± 13.47 66.58 ± 11.83
Accuracy 62.41 ± 4.7 57.79 ± 2.93 59.1 ± 5.64 60.04 ± 2.19 67.42 ± 1.83
SVM Sensitivity 61.19 ± 8.37 62.47 ± 4.98 61.47 ± 3.64 58.91 ± 5.45 62.19 ± 4.54
Specificity 63.77 ± 2.76 53.2 ± 5.07 58.48 ± 12.38 71.37 ± 4.14 72.68 ± 5.94
Accuracy 66.84 ± 3.39 62.49 ± 1.88 64.18 ± 4.34 62.22 ± 4.53 65.51 ± 3.81
LDA Sensitivity 68.11 ± 3.49 60.83 ± 3.81 67.95 ± 12.1 64.14 ± 3.44 66.99 ± 4.59
Specificity 65.56 ± 4.61 64.11 ± 2.09 60.38 ± 4.21 60.11 ± 10.47 64.13 ± 6.27
Accuracy 56.99 ± 2.88 61.36 ± 4.23 56.26 ± 6.83 66.59 ± 10.22 61.71 ± 3.3
KNN Sensitivity 51.46 ± 2.08 62.68 ± 9.13 62.99 ± 7.49 65.15 ± 10.38 62.35 ± 4.02
Specificity 62.47 ± 6.91 59.83 ± 7.16 49.25 ± 6.21 68.08 ± 10.5 61.33 ± 9.21

1 the nonlinear features. Results illustrated that the SVM clas- 30 rate of 68.26 ± 1.8% using the linear features. Table. 5b
2 sifier achieved the highest accuracy rate of 67.42 ± 1.83%. 31 illustrates the performance of Like/Dislike conditions identi-
3 In the algorithm, the most informative revealed extracted 32 fication based on the nonlinear features. In the algorithm, the
4 feature was the spectral entropy features, katz and higuchi 33 most informative revealed extracted features were sample
5 fractal dimension, ACS, AES, AC, AE, SDCS, and SDES 34 entropy, approximation entropy, Katz, and Higuchi fractal
6 from the combination of theta, alpha, beta, and gamma 35 dimensions from the delta, theta, beta, and gamma frequency
7 frequency bands in channels Fp2, F3, F8, C3, Cz, Pz, 36 bands in channels Fp1, F4, F8, C3, and O1. The RF classifier
8 and O1. The reason for higher outcome for the selected 37 achieved the accuracy rate of 65.16 ± 6.8% in comparison
9 nonlinear features is a wider scattering with less overlap 38 with the RF classifier using the nonlinear features. In this
10 in the feature space that this in turn could provide more 39 case, the chaotic features attain our objective in identifying
11 accurate result. As a consequence of this advantage, a lower 40 Like/Dislike task with significant accuracy result.
12 number of dimensions is required to be applied to obtain 41 Most of studies are based on the combination of genders
13 the best accuracy, which means faster and less complicated 42 in one group that the achieved results for the third group
14 processing. 43 in the our study are different from the previous studies.
15 Performance of male group classification: table. 4 illus- 44 Specifically, we found that the theta frequency band in cen-
16 trates the performance of Like/Dislike conditions identifi- 45 tral (C3) region was found the most discriminative channel.
17 cation based on the nonlinear features. In the algorithm, 46 Additionally, the linear features with the KNN classifier
18 the most informative revealed extracted features were the 47 reach the highest accuracy rate of 68.26 ± 1.8%. While,
19 sample entropy, approximate entropy and the Higuchi fractal 48 Golnar-Nik et al.] [10] showed that the accuracy rate for
20 dimension from the combination of delta, theta, alpha, and 49 identifying Like/Dislike preferences was 63.00%. In another
21 beta frequency bands in channels F3, F4, F8, P3, P4, and 50 study, Pereira et al. [47] showed that the frontal lobe were
22 Pz. The RF classifier achieved the best accuracy rate of 51 activated in a Like/Dislike music task, whilst in our study
23 71.33 ± 14.07% using nonlinear features. 52 the region was central (C3) region. Also, Andersen and
24 Performance of combination of male and female group: 53 Cui [48] showed that the frontal area and parietal regions
25 table. 5a illustrates the performance of Like/Dislike con- 54 were interconnected for several factors of decision-making.
26 ditions identification based on the linear features. In the 55 The incremental accuracy achievements using the nonlinear
27 algorithm, the most informative revealed extracted feature 56 features in the three groups could be attributable to the
28 was the PSD feature from the theta frequency band in 57 complexity of EEG signals.
29 channel C3. The KNN classifier achieved the best accuracy 58 In our study, we also considered the required average time
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

TABLE 4: Like/Dislike condition classification by using nonlinear features for male subjects.

Classifier Metric Delta Theta Alpha Beta All Bands

Accuracy 58.14 ± 6.93 53.49 ± 5.4 53.95 ± 6.18 52.09 ± 8.14 71.33 ± 14.07
RF Sensitivity 52.72 ± 12.3 65 ± 6.1 50.57 ± 10.36 48.82 ± 13.46 70.7 ± 9.08
Specificity 62.8 ± 9.1 60.01 ± 9.7 58.23 ± 7.05 57.09 ± 14.71 65.22 ± 10.39
Accuracy 66.44 ± 4.07 60.30 ± 4.68 52.94 ± 5.68 62.34 ± 3.57 63.12 ± 3.01
SVM Sensitivity 68.99 ± 10.72 59.17 ± 7.99 57.04 ± 12.96 71.33 ± 5.53 59.71 ± 2.58
Specificity 63.92 ± 7.65 65.39 ± 5.29 53.76 ± 5.72 59.27 ± 4.93 77.03 ± 4.89
Accuracy 63.64 ± 2.05 60.36 ± 2.06 59.72 ± 2.12 65.74 ± 6.07 63.65 ± 4.41
LDA Sensitivity 70.55 ± 3.97 59.61 ± 5.04 64.07 ± 3.69 73.52 ± 7.92 66.89 ± 7.54
Specificity 57.01 ± 3.56 61.08 ± 8.56 55.38 ± 2.06 57.97 ± 6.13 60.41 ± 4.28
Accuracy 61.71 ± 2.36 52.38 ± 5.83 52.41 ± 2.36 55.16 ± 3.95 66.49 ± 4.63
KNN Sensitivity 63.53 ± 4.21 49.72 ± 2.81 44.11 ± 3.49 58.11 ± 3.73 57.17 ± 1.72
Specificity 59.88 ± 3.48 55.04 ± 13.33 60.78 ± 6.81 52.15 ± 5.98 75.83 ± 9.04

TABLE 5: Like/Dislike condition classification by using (a) linear and (b) nonlinear features for combined male and female
subjects.
(a) Linear features for combined male and female subjects.
Classifier Metric Theta All Bands
Accuracy 62.25 ± 8.05 62.25 ± 8.05
RF Sensitivity 63.96 ± 14.08 63.96 ± 14.08
Specificity 60.14 ± 11.24 60.14 ± 11.24
Accuracy 57.29 ± 1.6 57.29 ± 1.6
SVM Sensitivity 53.33 ± 3 53.33 ± 3
Specificity 81.01 ± 0.75 81.01 ± 0.75
Accuracy 54.93 ± 2.7 54.93 ± 2.7
LDA Sensitivity 53.15 ± 4.41 53.15 ± 4.41
Specificity 66.55 ± 2.32 66.55 ± 2.32
Accuracy 68.26 ± 1.8 68.26 ± 1.8
KNN Sensitivity 64.02 ± 6.22 64.02 ± 6.22
Specificity 72.35 ± 6.46 72.35 ± 6.46

(b) Nonlinear features for combined male and female subjects.

Classifier Metric Delta Theta Beta Gamma All Bands
Accuracy 52.58 ± 6.98 50.32 ± 5.08 63.22 ± 5.93 50.32 ± 9.02 65.16 ± 6.78
RF Sensitivity 57.27 ± 15.53 51.53 ± 8.38 58.82 ± 9.83 53.95 ± 12.93 64.55 ± 10.12
Specificity 49.77 ± 13.63 49.96 ± 10.8 67.35 ± 8.21 50.22 ± 13.01 66.51 ± 14.64
Accuracy 60.21 ± 3.86 60.18 ± 2.84 62.52 ± 1.78 60.29 ± 3.09 61.52 ± 3.83
SVM Sensitivity 65.59 ± 3.22 67.52 ± 5.33 64.17 ± 6.81 45.04 ± 3.16 62.69 ± 2.38
Specificity 54.85 ± 5.21 52.91 ± 2.11 60.75 ± 5.78 75.65 ± 3.77 60.36 ± 5.56
Accuracy 59.83 ± 1.69 58.59 ± 2.29 59.78 ± 2.31 55.67 ± 1.53 57.4 ± 3.68
LDA Sensitivity 59.48 ± 3.47 66.07 ± 3.49 59.7 ± 5.14 61.74 ± 2.29 58.61 ± 2.43
Specificity 60.18 ± 1.27 51.29 ± 2.79 59.1 ± 8.91 49.69 ± 2.1 56.22 ± 5.24
Accuracy 52.9 ± 3.62 48.06 ± 4.74 51.28 ± 6.57 52.19 ± 2.91 56.21 ± 2.36
KNN Sensitivity 48.32 ± 3.23 46.96 ± 3.24 50.09 ± 7.01 50.98 ± 3.89 56.77 ± 2.16
Specificity 57.3 ± 9.04 49.1 ± 7.37 52.46 ± 6.8 53.4 ± 2.69 55.65 ± 2.81

1 for performing the Like/Dislike task that for the three groups 15 depend on linear and nonlinear features. According to the
2 it was approximately 2.5 seconds, which is in good agree- 16 female investigation achievements it is evident that the
3 ment with the studies conducted by Witchalls [49]. Addi- 17 most discriminative frequency for the females was the
4 tionally, We found Dislike decision-making happens a little 18 combination of theta, alpha, and beta frequency bands in
5 later than Like, more specifically the Like decision-making 19 frontal (Fp1, Fp2, F7, F8), central (Cz) , and parietal
6 took place in 2.2 ± 0.003 second and Dislike decision- 20 (Pz) regions. In the classification part, by making use of
7 making took place in 2.6 ± 0.02 second. Our experiments 21 the energy of wavelet coefficient feature with the SVM
8 showed that there are several human protocols for different 22 classifier the females group achieved the best accuracy of
9 types of subjects, which are missed in the computations. In 23 71.51 ± 5.10% for Like/Dislike identification. For male
10 future research, we will increase the number of subjects and 24 group, linear features extracted from the delta, theta, alpha,
11 improve the accuracy by using improving the mathematical 25 beta, and gamma frequency bands were insignificant in
12 methods of classification. 26 identifying Like/Dislike conditions, As Fig. 7 shows.

13 In summary, the identification of disctinctive frequency 27 According to the male group, it is evident that the most
14 bands, brain regions and Like/Dislike conditions are highly 28 discriminative frequency was the combination of delta, theta,
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

(a) Delta frequency band t-SNE projection. (b) Theta frequency band t-SNE projection.

(c) Alpha frequency band t-SNE projection. (d) Beta frequency band t-SNE projection.

(e) Gamma frequency band t-SNE projection.

FIGURE 7: Distribution of linear features in different frequency bands (a) Delta, (b) Theta, (C) Alpha, (d) Beta, and (e)
Gamma in a Like/Dislike task.

VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

1 alpha and beta bands in frontal (F3, F4, F8) and parietal (P3, 60 Electroencephalogram (EEG) of preferences,” Expert Systems with Appli-
2 P4, Pz) regions. In the classification part, by making use 61 cations, vol. 39, no. 16, pp. 12 378–12 388, 2012.
62 [7] R. N. Khushaba, C. Wise, S. Kodagoda, J. Louviere, B. E. Kahn, and
3 of the sample entropy, approximation entropy and Higuchi 63 C. Townsend, “Consumer neuroscience: Assessing the brain response to
4 fractal dimension features and RF classifier the male group 64 marketing stimuli using electroencephalogram (EEG) and eye tracking,”
5 reach the best accuracy of 71.33 ± 14.07% for Like/Dislike 65 Expert systems with applications, vol. 40, no. 9, pp. 3803–3812, 2013.
66 [8] B. Yilmaz, S. Korkmaz, D. B. Arslan, E. Güngör, and M. H. Asyalı,
6 identification. Finally, according to the the combined group, 67 “Like/dislike analysis using EEG: determination of most discrimina-
7 it is evident that the most discriminative frequency band was 68 tive channels and frequencies,” Computer methods and programs in
8 the theta in central (C3) region. In the classification part, 69 biomedicine, vol. 113, no. 2, pp. 705–713, 2014.
70 [9] T. Z. Ramsøy, M. Skov, M. K. Christensen, and C. Stahlhut, “Frontal brain
9 by making use of the linear features and KNN classifier the 71 asymmetry and willingness to pay,” Frontiers in neuroscience, vol. 12, p.
10 combined female and male groups reach the best accuracy 72 138, 2018.
11 of 68.26 ± 1.8% for Like/Dislike identification. 73 [10] P. Golnar-Nik, S. Farashi, and M.-S. Safari, “The application of EEG
74 power for the prediction and interpretation of consumer decision-making:
75 a neuromarketing study,” Physiology & behavior, vol. 207, pp. 90–98,
12 V. CONCLUSION 76 2019.
13 In the present study, a neuromarketing experimental task by 77 [11] M. Aldayel, M. Ykhlef, and A. Al-Nafjan, “Deep Learning for EEG-Based
Preference Classification in Neuromarketing,” Applied Sciences, vol. 10,
14 using the EEG amplifier was designed to display image of 78 79 no. 4, p. 1525, 2020.
15 shoe products for males and females. Brain regions related 80 [12] A. Hekmatmanesh, V. Zhidchenko, K. Kauranen, K. Siitonen, H. Han-
16 to the decision-making procedure of Like/Dislike products 81 droos, S. Soutukorva, and A. Kilpeläinen, “Biosignals in human factors
research for heavy equipment operators: A review of available methods
17 were considered for different genders. Therefore, different 82
83 and their feasibility in laboratory and ambulatory studies,” IEEE Access,
18 linear and nonlinear features were extracted from different 84 2021.
19 frequency bands and brain regions. The Wilcoxon Rank 85 [13] A. Hekmatmanesh, P. H. Nardelli, and H. Handroos, “Review of the state-
20 Sum statistical test was then applied on the features to 86 of-the-art of brain-controlled vehicles,” IEEE Access, 2021.
87 [14] S. G. H. Meyerding and C. M. Mehlhose, “Can neuromarketing add
21 specify distinctive frequencies and regions due to visual 88 value to the traditional marketing research? An exemplary experiment
22 stimulation task. Additionally, the best linear/nonlinear fea- 89 with functional near-infrared spectroscopy (fNIRS),” Journal of Business
23 tures for identifying like/dislike classes were considered 90 Research, vol. 107, pp. 172–185, 2020.
91 [15] L. H. Chew, J. Teo, and J. Mountstephens, “Aesthetic preference recogni-
24 with respect to the genders. The extracted features from the 92 tion of 3d shapes using eeg,” Cognitive neurodynamics, vol. 10, no. 2, pp.
25 distinctive EEG channels were categorized by using the RF, 93 165–173, 2016.
26 SVM, LDA, and KNN classifiers. Some part of significant 94 [16] M. Yadava, P. Kumar, R. Saini, P. P. Roy, and D. Prosad Dogra, “Analysis
95 of eeg signals and its application to neuromarketing,” Multimedia Tools
27 results showed that partially different areas of the brain 96 and Applications, vol. 76, no. 18, pp. 19 087–19 111, 2017.
28 were activated during Like/Dislike tasks in comparison with 97 [17] A. Hakim, S. Klorfeld, T. Sela, D. Friedman, M. Shabat-Simon, and
29 other studies for a mixed gender group. In addition, more 98 D. J. Levy, “Pathways to consumers’ minds: Using machine learning and
99 multiple eeg metrics to increase preference prediction above and beyond
30 areas and frequency ranges were activated in female’s brain 100 traditional measurements,” bioRxiv, p. 317073, 2018.
31 during shopping in comparison with male. The classification 101 [18] M. Alimardani and M. Kaba, “Deep learning for neuromarketing; classi-
32 results illustrated that the SVM classifier achieved more 102 fication of user preference using eeg signals,” in 12th Augmented Human
103 International Conference, 2021, pp. 1–7.
33 accurate results in comparison with the other classifier for 104 [19] F. R. Mashrur, K. M. Rahman, M. T. I. Miya, R. Vaidyanathan, S. F. Anwar,
34 female groups. Additionally, the combination of frequency 105 F. Sarker, and K. A. Mamun, “Bci-based consumers’ choice prediction
35 bands has potential of achieving more accurate results for 106 from eeg signals: An intelligent neuromarketing framework,” Frontiers in
107 Human Neuroscience, vol. 16, 2022.
36 identifying Like/Dislike (willing ti pay or not) conditions by 108 [20] C. A. E. Kothe and T.-P. Jung, “Artifact removal techniques with signal
37 means of a classifier in comparison with extracting features 109 reconstruction,” Apr. 28 2016, uS Patent App. 14/895,440.
38 from individual frequency bands with the same classifier. 110 [21] C.-Y. Chang, S.-H. Hsu, L. Pion-Tonachini, and T.-P. Jung, “Evaluation
111 of artifact subspace reconstruction for automatic EEG artifact removal,” in
112 2018 40th Annual International Conference of the IEEE Engineering in
39 REFERENCES 113 Medicine and Biology Society (EMBC). IEEE, 2018, pp. 1242–1245.
40 [1] B. Veronica, “Brief history of neuromarketing,” J Bert Rus, pp. 119–121, 114 [22] S. Makeig, T.-P. Jung, A. J. Bell, D. Ghahremani, and T. J. Sejnowski,
41 2009. 115 “Blind separation of auditory event-related brain responses into indepen-
42 [2] B. Y. Ozkara and R. Bagozzi, “The use of event related potentials brain 116 dent components,” Proceedings of the National Academy of Sciences,
43 methods in the study of conscious and unconscious consumer decision 117 vol. 94, no. 20, pp. 10 979–10 984, 1997.
44 making processes,” Journal of Retailing and Consumer Services, vol. 58, 118 [23] A. J. Bell and T. J. Sejnowski, “An information-maximization approach
45 p. 102202, 2021. 119 to blind separation and blind deconvolution,” Neural computation, vol. 7,
46 [3] M. D. Bercea, “Anatomy of methodologies for measuring consumer be- 120 no. 6, pp. 1129–1159, 1995.
47 havior in neuromarketing research,” in Proceedings of the LCBR European 121 [24] A. Hyvärinen and E. Oja, “Independent component analysis: algorithms
48 Marketing Conference, 2012, pp. 1–14. 122 and applications,” Neural networks, vol. 13, no. 4-5, pp. 411–430, 2000.
49 [4] C. Solnais, J. Andreu-Perez, J. Sánchez-Fernández, and J. Andréu-Abela, 123 [25] ucsd., “Iclabel tutorial: EEG independent component labeling,” Retrieved
50 “The contribution of neuroscience to consumer research: A concep- 124 from https://labeling.ucsd.edu/tutorial/labels, 2021, accessed January 3,
51 tual framework and empirical review,” Journal of economic psychology, 125 2021.
52 vol. 36, pp. 68–81, 2013. 126 [26] A. Hekmatmanesh, H. Wu, A. Motie-Nasrabadi, M. Li, and H. Handroos,
53 [5] G. Vecchiato, J. Toppi, L. Astolfi, F. D. V. Fallani, F. Cincotti, D. Mattia, 127 “Combination of discrete wavelet packet transform with detrended fluc-
54 F. Bez, and F. Babiloni, “Spectral EEG frontal asymmetries correlate with 128 tuation analysis using customized mother wavelet with the aim of an
55 the experienced pleasantness of TV commercial advertisements,” Medical 129 imagery-motor control interface for an exoskeleton,” Multimedia Tools
56 and biological engineering and computing, vol. 49, no. 5, pp. 579–583, 130 and Applications, vol. 78, no. 21, pp. 30 503–30 522, 2019.
57 2011. 131 [27] A. Hekmatmanesh, H. Wu, M. Li, A. M. Nasrabadi, and H. Handroos,
58 [6] R. N. Khushaba, L. Greenacre, S. Kodagoda, J. Louviere, S. Burke, 132 “Optimized mother wavelet in a combination of wavelet packet with de-
59 and G. Dissanayake, “Choice modeling and the brain: A study on the 133 trended fluctuation analysis for controlling a remote vehicle with imagery

12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2022.3201488

1 movement: A brain computer interface study,” in New Trends in Medical 73 ATEFE HASSANI received her B.Sc. degree in
2 and Service Robotics. Springer, 2019, pp. 186–195. 74 Electrical Engineering from the University of
3 [28] D. Q. Phung, D. Tran, W. Ma, P. Nguyen, and T. Pham, “Using Shannon 75 Zanjan, Iran, in 2017, and the M.Sc. degree in
4 Entropy as EEG Signal Feature for Fast Person Identification.” in ESANN, 76 Biomedical Engineering from Shahed University,
5 vol. 4, no. 1, 2014, pp. 413–418. 77 Tehran, Iran, in 2020. She is currently a visiting
6 [29] A. Hekmatmanesh, R. M. Asl, H. Wu, and H. Handroos, “Eeg control of a 78 researcher in the Bio-Imaging lab, University of
7 bionic hand with imagination based on chaotic approximation of largest 79 Antwerp, Belgium. Her current research interests
8 lyapunov exponent: A single trial bci application study,” IEEE Access,
80 include biomedical signal processing, machine
9 vol. 7, pp. 105 041–105 053, 2019.
81 learning, deep learning, and computational neu-
10 [30] N. Kannathal, M. L. Choo, U. R. Acharya, and P. K. Sadasivan, “Entropies
11 for detection of epilepsy in EEG,” Computer methods and programs in 82 roscience.
12 biomedicine, vol. 80, no. 3, pp. 187–194, 2005.
13 [31] A. Hekmatmanesh, M. Mikaeili, K. Sadeghniiat-Haghighi, H. Wu, H. Han-
14 droos, R. Martinek, and H. Nazeran, “Sleep spindle detection and pre-
15 diction using a mixture of time series and chaotic features,” Advances in
16 Electrical and Electronic Engineering, vol. 15, no. 3, 2017.
17 [32] H. Niknazar, A. M. Nasrabadi, and M. B. Shamsollahi, “Volumetric
18 behavior quantification to characterize trajectory in phase space,” Chaos,
19 Solitons & Fractals, vol. 103, pp. 294–306, 2017.
20 [33] F. Takens, “Detecting strange attractors in turbulence,” in Dynamical
21 systems and turbulence, Warwick 1980. Springer, 1981, pp. 366–381.
22 [34] N. H. Packard, J. P. Crutchfield, J. D. Farmer, and R. S. Shaw, “Geometry
23 from a time series,” Physical review letters, vol. 45, no. 9, p. 712, 1980. 83 AMIN HEKMATMANESH received his bache-
24 [35] A. Hekmatmanesh, R. M. Asl, H. Handroos, and H. Wu, “Optimizing 84 lor’s in Electrical Engineering from Science and
25 largest lyapunov exponent utilizing an intelligent water drop algorithm: A 85 Research of Fars University, Shiraz, Iran, 2010.
26 brain computer interface study,” in 2019 5th International Conference on 86 His Master’s degree was in Biomedical Engi-
27 Event-Based Control, Communication, and Signal Processing (EBCCSP). 87 neering from Shahed University in Tehran, Iran,
28 IEEE, 2019, pp. 1–5. 88 2013. His Master’s thesis was about analyzing
29 [36] R. Genuer, J.-M. Poggi, and C. Tuleau-Malot, “Variable selection using 89 Sleep EEG signal processing, learning and nega-
30 random forests,” Pattern recognition letters, vol. 31, no. 14, pp. 2225– 90 tive emotional memory. He received his PhD on
31 2236, 2010.
91 brain-controlled ankle foot and hand orthosis and
32 [37] W. Lin, Z. Wu, L. Lin, A. Wen, and J. Li, “An ensemble random forest
33 algorithm for insurance big data analysis,” IEEE Access, vol. 5, pp.
92 mobile vehicle robots using the EEG in the Lab-
34 16 568–16 575, 2017. 93 oratory of Intelligent Machines in Lappeenranta University of Technology,
35 [38] V. N. Vapnik, “Methods of pattern recognition,” in The nature of statistical 94 2016-2019. Since 2020 he is working as a Post-Doc on heavy machine
36 learning theory. Springer, 2000, pp. 123–180. 95 operator’s health monitoring and signal processing for horse simulators
37 [39] V. Vapnik, The nature of statistical learning theory. Springer science & 96 in the Laboratory of Intelligent Machines in Lappeenranta University of
38 business media, 2013. 97 Technology.
39 [40] A. B. Goldberg and X. Zhu, “New directions in semi-supervised learning,”
40 Ph.D. dissertation, University of Wisconsin–Madison, 2010.
41 [41] A. Hekmatmanesh, H. Wu, F. Jamaloo, M. Li, and H. Handroos, “A
42 combination of csp-based method with soft margin svm classifier and
43 generalized rbf kernel for imagery-based brain computer interface appli-
44 cations,” Multimedia Tools and Applications, vol. 79, no. 25, pp. 17 521–
45 17 549, 2020.
46 [42] A. Hekmatmanesh, F. Jamaloo, H. Wu, H. Handroos, and A. Kilpeläinen,
47 “Common spatial pattern combined with kernel linear discriminate and
48 generalized radial basis function for motor imagery-based brain computer
49 interface applications,” in AIP Conference Proceedings, vol. 1956, no. 1.
50 AIP Publishing LLC, 2018, p. 020003. 98 ALI MOTIE NASRABADI received the B.Sc.
51 [43] X. Feng, J. Yang, F. Luo, J. Chen, and X. Zhong, “Automatic modulation 99 degree in electronic engineering and the M.Sc.
52 recognition by support vector machines using wavelet kernel,” in J. Phy. 100 and Ph.D. degrees in biomedical engineering from
53 Conf. Series, vol. 48, no. 1, 2006, pp. 1264–126.
101 the Amirkabir University of Technology, Tehran,
54 [44] C. J. C. Burges, “A tutorial on support vector machines for pattern
102 Iran,in 1994, 1999, and 2004, respectively. In
55 recognition,” Data mining and knowledge discovery, vol. 2, no. 2, pp. 121–
56 167, 1998. 103 2004,he joined the Biomedical Engineering De-
57 [45] S. M. R. Noori, A. Hekmatmanesh, M. Mikaeili, and K. Sadeghniiat- 104 partment, Faculty of Engineering, Shahed Univer-
58 Haghighi, “K-complex identification in sleep eeg using melm-grbf clas- 105 sity, where he was an Assistant Professor, from
59 sifier,” in 2014 21th Iranian conference on biomedical engineering 106 2004 to 2011, an Associate Professor, from 2011
60 (ICBME). IEEE, 2014, pp. 119–123. 107 to 2017, and has been a Full Professor, since
61 [46] M.-P. Hosseini, A. Hosseini, and K. Ahi, “A review on machine learning 108 2017. He is currently a Scientific Advisor with the National Brain Mapping
62 for eeg signal processing in bioengineering,” IEEE reviews in biomedical 109 Laboratory, University of Tehran, Iran. His current research interests
63 engineering, vol. 14, pp. 204–218, 2020. 110 include brain–computer interfaces, biomedical signal processing, machine
64 [47] C. S. Pereira, J. Teixeira, P. Figueiredo, J. Xavier, S. L. Castro, and 111 learning, deep learning, nonlinear time series analysis, and computational
65 E. Brattico, “Music and emotions in the brain: familiarity matters,” PloS 112 neuroscience. He is a Board Member of the Iranian Society for Biomedical
66 one, vol. 6, no. 11, p. e27241, 2011. 113 Engineering and has served on the scientific committees for several national
67 [48] R. A. Andersen and H. Cui, “Intention, action planning, and decision 114 conferences and review boards of five scientific journals.
68 making in parietal-frontal circuits,” Neuron, vol. 63, no. 5, pp. 568–583,
69 2009.
70 [49] newsweek., “Pushing the buy button,” Retrieved from https://www.
71 newsweek.com/pushing-buy-button-123737, 2021, accessed January 3,
72 2021.

VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/