Discrimination of Customers Decision-Making in a Like/Dislike Shopping Activity Based on Genders: A Neuromarketing Study

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, 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 nonlinear 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.


I. INTRODUCTION
Traditionally, marketing methods (newspaper advertisements 23 and television commercials) were invested based on the cus- 24 tomer's spoken information for identifying their interests. 25 In some points, the investors got success but in many cases 26 investments failed. 27 The associate editor coordinating the review of this manuscript and approving it for publication was Xinyu Du .
Neuromarketing is a topic for marketing research area 28 that uses neuroscience-related techniques to study consumers 29 behavior. The concept of neuromarketing was first introduced 30 by psychologist [1] in the 1990s at Harvard University. It is 31 revealed that the decision-making process about unknown 32 brands happens on an unconscious level of mind [2]. 33 Neuromarketing techniques such as electroencephalography 34 (EEG) signal processing has been employed and results 35 are attributed to the human unconscious information during 36 prefrontal gamma asymmetry was related to willing to pay 93 responses significantly. Then, Golnar-Nik et al. [10] consid-94 ered the hypothesis if the PSD feature from the EEG is a suit-95 able approach for predicting the customers decision-making 96 in a Like/Dislike task. In the paper the features were clas-97 sified by using Support Vector Machine (SVM) and Linear 98 Discriminant Analysis (LDA). Additionally, authors consid-99 ered the customers preferences in an advertisement task with 100 respect to the background colors of products. Statistical anal-101 ysis based on the PSD showed the same results as in [9]. 102 In summary, the obtained locations relative to the willing 103 to buying for a product (Like/Dislike tasks) were Centro-104 parietal locations, namely Fp1, Cp3 and Cpz. In addition, 105 significant changes were observed in the frontal electrodes, 106 namely F4 and Ft8. 107 Aldayel et al. [11] considered customers preferences using 108 EEG signals. In the algorithm, the PSD and valence features 109 were extracted from the filtered EEG signals and then fed 110 into four different classifiers, named deep learning, SVM, 111 Random Forest (RF), and k-Nearest Neighbor (KNN), that 112 the deep learning provided the best accuracy and precision 113 results. Additionally, results showed that the RF classifier 114 obtained similar results as in the deep learning method. The 115 advantage of the deep learning method is the higher potential 116 of producing more accurate and precise results for multi-class 117 identifications problems than the RF and SVM classifiers. 118 Also, deep learning has capability of handling a large number 119 of data, which has been applied on a wide range of different 120 biosignal processing in health monitoring and brain computer 121 interface fields [12], [13]. The disadvantage of the deep learn-122 ing algorithm is a large number of input data is required for 123 the training phase. Additionally, the training phase is very 124 time consuming process. 125 In another study, Meyerding et al. [14] focused on the 126 neuronal activation of brain during decision-making of label 127 brands for different products. In the algorithm, the difference 128 of neurons activities in the prefrontal cortex (PFC) were 129 measured during subjects coping with two labels brands. 130 The achievements based on the fNIRS showed that the PFC 131 activity for individual subjects increased significantly. The 132 main limitation of the study was participating a low number 133 of subjects in the experiment.

134
Most recent neuromarketing studies focused on the com-135 bination of subjects, while gender is an important parameter 136 which has not been taken into account in the investigations. 137 The primary functions of human's perception is various for 138 different genders points of view and it would be an impressive 139 parameter for improving the neuromarketing models, which 140 has not been considered yet. In addition, linear features are 141 exhaustively studied in the computational parts and a lack 142 of nonlinear feature investigation exists. Table. 1 present a 143 review of some experimental neuromarketing studies.

144
The first contribution of the present study is identifying the 145 most discriminative areas of brain in a Like/Dislike (willing 146 to buy) task with respect to the subject's gender as an impor-147 tant factor. The second contribution is considering the EEG 148 VOLUME 10, 2022 Section IV is the results and discussion; and section V is the 159 conclusion.

161
The EEG signals recorded from twenty (10 males and 162 10 females) healthy and graduated right-handed students. 163 Before the experiment, the required regulations and condi-164 tions for the task were explained to the subjects and a consent  and the main question for women was are you willing to pay 185 for the shoes for herself and male subjects should answer to 186 the question are you willing to pay for the shoes for your 187 life partner (wife). Paradigm blocks were used to display 188 images in the MATLAB software simulation environment. 189 The structure of experimental protocol is shown in Fig. 1.

191
A 16-channel EEG amplifier (g.USBamp, g.tec, Austria) was 192 used to record electrical potential. The EEG electrodes were 193 installed based on the international 10-20 electrode location 194 system, in which enable us to cover almost all areas of a head. 195 The main purpose of the 10-20 standard is providing elec-196 trode instalment using a small number of electrodes (typically 197 21 electrodes) for recording EEG. In our experiment, the right 198 ear and Fpz electrodes were set as the GROUND (GND) and 199 common reference for all the channels, respectively. Our elec-200 trode instalment is showed in Fig. 2. The EEG was recorded 201 with the sampling frequency of 256 Hz, and a High-pass filter 202 with a cut off frequency of 0.1 Hz and a notch filter with a 203 cut off frequency of 50 Hz were applied. Table. 2 shows the 204 number of decisions for subjects. After collecting the EEG 205 data, the preprocessing method is applied to remove noise.  In the second noise removal step, an ICA algorithm applied 233 on the denoised signal. The basic hypothesis concept of ICA 234 is the EEG signal is weighted by using linear combination 235 of electrical potentials, which are generated from indepen-236 dent brain sources [22], [23]. The goal of ICA algorithm 237 is finding the least Gaussian state in a new space in which 238 leads to the identification of the main sources. The ICA is not 239 able to separate the sources of noises if they are completely 240 Gaussian. If the computed components were not independent, 241 then the ICA seek for the most independent space to separate 242 linear sources [24]. Therefore, we mapped the EEG data 243 from the sensor space into the source space and removed the 244 major interference sources including eye components, muscle 245 components, heart components, as well as line noise, and 246 channel noise [25].

248
To identify the active brain regions in our marketing stimu-249 lus experiments, different linear and nonlinear features from 250 five different frequency bands were extracted and analyzed. 251 These analysis were applied on three groups with respect to 252 the genders (male, female, and a combination of male and 253 female). To this end, two sets of linear and nonlinear features 254 were extracted from the EEG signals. 255

256
The linear feature extraction approaches include the energy 257 of wavelet coefficients and PSD. To this end, the EEG 258 signals were decomposed by using the Discrete Wavelet 259 Transform (DWT) algorithm to obtain different frequency 260 bands. Therefore, a mother wavelet is set in the DWT, 261 named Daubechies 8 (db8) and then the DWT was applied 262 for five levels to reach the aim frequency bands [26], 263 [27]. The energy of DWT coefficients were computed as 264 follows: The PSD feature computed by means of modified peri-269 odogram algorithm, which is a non-parametric estimation. 270 The PSD algorithm involves following steps: 1) multiplying 271 the input time series with a non-rectangular window such as 272 Hamming; 2) applying the Fourier transform function; and 3) 273 computing the density of power spectrum estimation by using 274 the Fourier transform size. Typically, the non-parametric 275 methods have less computational complexity than the para-276 metric models. Therefore, the PSD features were extracted 277 from the frequency bands consist of Delta (0.5-4 Hz), Theta 278 (4-8 Hz), Alpha (8-13 Hz), Beta (13-30 Hz), and Gamma 279 (30)(31)(32)(33)(34)(35)(36)(37)(38)(39)(40). In the computations, there was no restriction for 280 selecting the frequency bands.

282
Brain function is a non-stationary system that produces sig- and fractal dimensions [26].

291
The idea of entropy was first proposed in the thermo-292 dynamic field to measure the trajectory of a system. The 293 entropy concept describes the behavior of a part of a trajectory 294 that can be predicted from the rest. In the computations, 295 the higher entropy in a system leads to higher complex-296 ity, which means less chance for predicting the systems 297 behaviour [30], [31]. In the present study, we have employed Katz fractal dimensions were employed [31]. According to the Taken's theory [33], growing the tra-316 jectory of EEG signals in a reconstructed phase space is 317 applicable using the time series X (t) = {x 1 , x 2 , . . . , x N } and 318 two delayed input signals as follows [34], [35]: (2) 321 where µ and τ are embedding dimension and delay, 322 respectively, which are obtained based on the false 323 nearest-neighbors and mutual information methods. To extract 324 features, the Euclidean distance matrices (T ) of trajectories 325 (L) with length N between all the state vectors are computed 326 follows: Matrix T is obtained by removing the main diagonal of 329 matrix T . The last column of the matrix T is then removed 330 and shifted to the left. Matrix T with dimension (M × (M − 331 2)) represents the difference between the distances of state 332 vectors and trajectory motion.  SVM is a supervised learning classifier, which is based on 402 statistical learning theory. The first nonlinear SVM model 403 introduced by Vapnik [38]. Conceptually, the SVM model 404 is a hyperplane or line that separates a set of positive and 405 negative samples, namely support vectors, by using a max-406 imum distance [39], [40]. Recently, several modifications 407 of the SVM has been developed to increase the accuracy 408 and precision results [41], [42]. In the computations, the 409 boundaries of the two classes may not be linearly separable. 410 Therefore, features are mapped from the input space to a 411 feature space with higher dimension by means of nonlinear 412 kernels. The mapped feature space dimension is increased 413 until the features are separable linearly [43]. Different types 414 of kernels has been employed for the SVM decision func-415 tion, the most common used kernel in the SVM algorithm 416 is the Radial Basis Function (RBF). The advantages of 417 RBF are using Gaussian shape for the distributions and 418 enabling a feature space with unlimited dimensions [44]. 419 Also, several method developed to improve the capability of 420 RBF [41], [45].

422
LDA is a linear model commonly used for supervised clas-423 sification problems and dimensionality reduction. This tech-424 nique offer a good separation between different classes and 425 avoid overfitting. As a result, computational costs will be 426 significantly reduced and classification will be more accurate 427 by projecting the given n-dimensional feature space onto a 428 smaller feature space while maximizing the class separabil-429 ity [46].

431
KNN is a non-parametric and supervised learning algorithm. 432 This algorithm is simple to implement and robust to the noisy 433 training data which can be used for both classification and 434 regression. In KNN, inputs are classified based on their K 435 neighbors. The disadvantage of the algorithm is the value 436 of K will always need to be determined, which may be 437 complicated. Since the distance between the data points for 438 all the samples of training dataset must be calculated, the 439 computation cost will be high. In the next step, the obtained 440 results are presented and discussed.

442
In the present study, 20 subjects participated in the exper-443 iment and EEG signals were recorded by using visual 444 Like/Dislike stimulation task. The above-mentioned algo-445 rithms were then applied on the EEG data to find the dis-446 tinctive frequency bands and the relative areas of brain which 447 were affected by willing to pay decision for a pair of shoes. 448 Here, the results of frequency bands, brain regions and the 449 effects of nonlinear and linear features on the distinguished 450 VOLUME 10, 2022     To identify the discriminated frequency bands between 542 Like/Dislike groups, well-established RF, SVM, LDA, and 543 KNN classifiers were employed. In the process, a matrix 544 size of 15 × 14 × 160 was formed for individual subjects, 545 i.e., 15 channels, 14 features, and 16 Like/Dislike epochs. 546 In our study, the reported accuracy results for the Like/Dislike 547 classification are based on a 10-fold cross-validation. The 548 explanations are performed based on the genders.

549
Performance of female group classification: Table. 3a illus-550 trates the performance of the Like/Dislike conditions identi-551 fication based on the linear features. According to the RF, 552 SVM, LDA, and KNN classifiers results, the SVM classifier 553 achieved the highest accuracy of 71.51 ± 5.10%. In the 554 algorithm, the most informative revealed extracted feature 555 was the energy of wavelet coefficient from the combination 556

561
In the algorithm, the most informative revealed extracted 562 features were the spectral entropy features, katz, higuchi 563 fractal dimension, ACS, AES, AC, AE, SDCS, and SDES 564 from the combination of theta, alpha, beta, and gamma fre-565 quency bands in channels Fp2, F3, F8, C3, Cz, Pz, and 566 O1. The reason for higher outcome for the selected nonlin-567 ear features is a wider scattering with less overlap in the 568 feature space that this in turn could provide more accu-569 rate result. As a consequence of this advantage, a lower 570 number of dimensions is required to be applied to obtain 571 the best accuracy, which means faster and less complicated 572 processing.

573
Performance of male group classification: Table. 4   In our study, we also considered the required average time 616 for performing the Like/Dislike task that for the three groups

628
In summary, the identification of disctinctive frequency 629 bands, brain regions and Like/Dislike conditions are highly 630 depend on linear and nonlinear features. According to the 631 female investigation achievements it is evident that the most 632 discriminative frequency for the females was the combination 633 of theta, alpha, and beta frequency bands in frontal (Fp1, 634 Fp2, F7, F8), central (Cz), and parietal (Pz) regions. In the 635 classification part, by making use of the energy of wavelet 636 coefficient feature with the SVM classifier the females group 637 achieved the best accuracy of 71.51 ± 5.10% for Like/Dislike 638 identification. For male group, linear features extracted from 639 the delta, theta, alpha, beta, and gamma frequency bands were 640 insignificant in identifying Like/Dislike conditions, as Fig. 7 641 shows.

642
According to the male group, it is evident that the most 643 discriminative frequency was the combination of delta, theta, 644 alpha and beta bands in frontal (F3, F4, F8) and parietal (P3, 645 P4, Pz) regions. In the classification part, by making use 646 of the sample entropy, approximation entropy and Higuchi 647 fractal dimension features and RF classifier the male group 648 reach the best accuracy of 71.33 ± 14.07% for Like/Dislike 649 identification. Finally, according to the the combined group, 650 it is evident that the most discriminative frequency band was 651 the theta in central (C3) region. In the classification part, 652 by making use of the linear features and KNN classifier the 653 combined female and male groups reach the best accuracy of 654 68.26 ± 1.8% for Like/Dislike identification.

656
In the present study, a neuromarketing experimental task by 657 using the EEG amplifier was designed to display image of 658 shoe products for males and females. Brain regions related 659 to the decision-making procedure of Like/Dislike products 660 were considered for different genders. Therefore, different 661 linear and nonlinear features were extracted from different 662 frequency bands and brain regions. The Wilcoxon Rank Sum 663 statistical test was then applied on the features to specify 664 distinctive frequencies and regions due to visual stimulation 665 task. Additionally, the best linear/nonlinear features for iden-666 tifying Like/Dislike classes were considered with respect to 667 the genders. The extracted features from the distinctive EEG 668 channels were categorized by using the RF, SVM, LDA, and 669 KNN classifiers. Some part of significant results showed that 670 partially different areas of the brain were activated during 671 Like/Dislike tasks in comparison with other studies for a 672 mixed gender group. In addition, more areas and frequency 673 ranges were activated in female's brain during shopping in 674 comparison with male. The classification results illustrated 675 that the SVM classifier achieved more accurate results in 676 comparison with the other classifier for female groups. Addi-677 tionally, the combination of frequency bands has potential of 678 achieving more accurate results for identifying Like/Dislike 679 (willing to pay or not) conditions by means of a classifier 680

855
AMIN HEKMATMANESH received the bache-856 lor's degree in electrical engineering from the Sci-857 ence and Research of Fars University, Shiraz, Iran, 858 2010, the master's degree in biomedical engineer-859 ing from Shahed University, Tehran, Iran, in 2013, 860 and the Ph.D. degree in brain-controlled ankle 861 foot and hand orthosis and mobile vehicle robots 862 using the EEG from the Laboratory of Intelligent 863 Machines, Lappeenranta University of Technol-864 ogy, in 2019. His master's thesis was about analyz-865 ing sleep EEG signal processing, learning and negative emotional memory. 866 Since 2020, he has been working as a Postdoctoral Researcher on heavy 867 machine operator's health monitoring and signal processing for horse simu-868 lators at the Laboratory of Intelligent Machines, Lappeenranta University of 869 Technology.