Classification of Emotional Signals from the DEAP Dataset

Giuseppe Placidi

, Paolo Di Giamberardino

, Andrea Petracca

, Matteo Spezialetti

and Daniela Iacoviello

VI_Lab, c/o Department of Life, Health and Environmental Sciences, University of L’Aquila,

Via Vetoio, 67100, L’Aquila, Italy

Department of Computer, Control and Management Engineering Antonio Ruberti, Sapienza University of Rome,

Via Ariosto 25, 00185, Rome, Italy

Keywords: BCI, Emotional Signals, DEAP Dataset, Machine Learning, PCA, SVM.

Abstract: A Brain Computer Interface (BCI) is a useful instrument to support human communication, frequently

implemented by using electroencephalography (EEG). Regarding the used communication paradigm, a very

large number of strategies exist and, recently, self-induced emotions have been introduced. However, in

general the actual emotion-based BCIs are just binary, since they are capable of recognizing just a single

emotion. A crucial node is the introduction of more than a single emotional state for improving the

efficiency of a BCI. In order to be used in BCIs, signals from different emotional states have to be collected,

recognized and classified. In the present paper, a method for mapping several emotional states was

described and tested on EEG signals collected from a publicly available dataset for emotion analysis using

physiological signals (DEAP). The proposed method, its experimental protocol, and preliminary numerical

results on three different emotional states were presented and discussed. The method, based on multiple

binary classification, was capable of optimizing the most discriminative channels and the features

combination for each emotional state and of recognizing between several emotional states through a polling

system.

1 INTRODUCTION

BCIs provide new channels of output for the brain,

(Shih, 2012), yielding an enormous help to disabled

people (Shih, 2012; Kubler et al., 2005; Hochberg et

al., 2006). The neural activity is often recorded by

EEG (Fisch, 1999) and is based on event-related

signals (Babiloni et al., 2000). Recently, in (Placidi

et al., 2015a) an EEG-based BCI that used the

stimulus generated by the disgust produced by

remembering an unpleasant odor has been proposed

and it has been demonstrated to be particularly

useful for severely disabled people (Pistoia et al.,

2015).

Being the signals resulting by a self-induced

emotion weak, a series of competitive classification

strategies have been proposed (Placidi et al., 2015b;

Iacoviello et al., 2015a; Iacoviello et al., 2015b;

Iacoviello et al., 2015c). However, the BCI obtained

by using just “disgust” versus “relax” was only

binary, (Placidi et al., 2015c). In order to improve

Figure 1: Electrodes used in the 10-20 international brain

positioning system.

the efficiency of an emotional BCI (i.e. to increase

the cardinality of the “alphabet” and to reduce the

time necessary for communication), a series of

Placidi, G., Giamberardino, P., Petracca, A., Spezialetti, M. and Iacoviello, D.

Classiﬁcation of Emotional Signals from the DEAP dataset.

DOI: 10.5220/0006043400150021

In Proceedings of the 4th International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX 2016), pages 15-21

ISBN: 978-989-758-204-2

different emotions should be recognized and used

(Guler and Ubeyli, 2007). Before using different

emotions on a BCI, their characteristic expressions

(activation sites and specific features) have to be

discovered and compared. In the present paper, the

classification strategy proposed in (Iacoviello et al.,

2015c) is used on EEG signals collected in the

DEAP dataset (Koelstra et al., 2012), a database

containing a collection of physiological EEG signals

of emotions from different subjects both for negative

and positive emotions. In particular the participants

watched music videos and rated each video in terms

of arousal, valence, like/dislike, dominance, and

familiarity. As the subjects watched the videos, their

EEG and physiological signals were recorded. The

stimuli used in the experiment were selected in

different steps: first, 120 initial stimuli were

selected; then, a one-minute highlight part was

determined for each stimulus; finally, through a

web-based subjective assessment experiment, 40

final stimuli were selected. Being DEAP a reference

database for tagged EEG emotional signals freely

usable, we selected some of the stored experiments

in order to study the brain activations due both to

negative and positive emotions and to recognize the

most significant. In particular, goals of this paper

are: a) to verify that, for a subset of subjects from

the DEAP dataset, the activated brain region for a

negative emotion (negative valence and high

arousal) is located in the right brain hemisphere; b)

to classify positive emotions (high valence and high

arousal) from the selected subjects; c) to verify the

separation, in terms of activated channels and

selected features, between negative and positive

patterns; d) to propose a method for classifying

several emotional states to be used in future multi-

emotional BCI. The paper is organized as follows. In

Section II, the DEAP dataset and the experimental

protocol adopted are described along with the

considered classification method. In Section III the

obtained results are proposed and discussed, whereas

in Section IV the conclusions and future works are

outlined.

2 MATERIALS AND METHODS

The DEAP database consists of the EEG

physiological signals of 32 participants (16 men and

16 women, aged between 19 and 37, average: 26.9)

recorded while watching 40 one-minute long music

videos on different arguments. Before starting the

viewing, a two-minutes long EEG signal was

collected by each subject while relaxing watching a

fixation cross on the screen. The EEG signals,

sampled at 512 Hz, were recorded from the follo-

wing 32 positions (according to the international 10-

20 positioning system, see Figure 1): Fp1, AF3, F3,

F7, FC5, FC1, C3, T7, CP5, CP1, P3, P7, PO3, O1,

Oz, Pz, Fp2, AF4, Fz, F4, F8, FC6, FC2, Cz, C4, T8,

CP6, CP2, P4, P8, PO4, and O2. The proposed

music videos were demonstrated to induce emotions

to different users (Koelstra et al., 2012) represented

in the valence-arousal scale (Russell, 1980). The

participants had to rate each video in terms of

arousal, valence, like/dislike, dominance and

familiarity (the degree of valence and arousal was

ranged by using the self-assessment manikins

questionnaire). The same videos had an on-line

evaluation that could be used for comparison. The

videos were the same for all the participants but the

sequence of visualization for each subject was

random. As a first step, in the present study just the

dimensions valence and arousal were considered.

Data were provided both as they were acquired (raw

data) and in the preprocessed form. In this study, the

raw data were used and, before their usage, they

were filtered between 1 Hz and 46 Hz.

2.1 The Experimental Protocol

The main goal of this study was to use DEAP to map

the emotions through the EEG signals from different

subjects, by considering the results on the

classification of a strong negative emotion, the

disgust (Placidi et al., 2015b). To this aim, we

started by selecting subjects that experienced the

“strongest” and reciprocally “farthest” couple of

emotions, one corresponding to minimum negative

valence and maximum arousal (in the following

indicated with NVHA) and the other corresponding

to the maximum valence and maximum arousal (in

the following indicated with HVHA). Between the

selected subjects, we further selected those whose

self-assessment of NVHA and HVHA corresponded

to videos having the same on-line evaluation: this

was done in order to eliminate careless subjects

(possible cases of wrong evaluations). From the

selected subjects, besides the EEG signals corres-

ponding to these two emotional states, we extracted

the EEG signals corresponding to the relaxing phase.

In fact, after the selection of the subjects and of the

signals of the chosen emotions, we aimed at

classifying these two emotional states both with the

corresponding relaxing signals and reciprocally. The

one-minute signals corresponding to the emotional

state elicited by a music video was broken into non-

overlapping trials, 3.52 seconds long, and separately

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

used for classification. In the same way, also the

relaxing signals was divided in contiguous trials of

3.52 seconds.

2.2 The Classification Method

The method introduced in (Iacoviello et al., 2015a;

Iacoviello et al., 2015c) aimed at the classification of

EEG signals induced by remembering the disgust

produced by unpleasant odor: it was a self-induction,

without any external stimulation. One of the

conclusions of that research was that the most

involved channels were the T8 and P4 ones, both

belonging to the right brain hemisphere. In the

present study, the same method was used to classify

trials from two classes at once corresponding,

respectively, to NVHA (E1) versus Relax (R),

HVHA (E2) versus Relax (R) and E1 versus E2,

thus implicitly allowing a multiclass classification

(through the construction of a polling system). The

signals herein considered were not self-induced and

corresponded to the disgust elicited while watching

music videos. Data from the DEAP dataset were

produced by an external stimulation and, since the

stimulation involved different aspects (videos,

music, secondary emotions, and so on) they

guaranteed low specificity, from the localization

point of view.

Each acquired signal g (trial) was first processed by

diadic Wavelet decomposition to extract just the

useful information content (Daubechies, 1992;

Mamun et al., 2013). To this aim the Meyer wavelet

was used to provide the wavelet

transform:

(k2

− j

) = 2

j /2

g(n)

k − n

()



(1)

thus decomposing the signal in the approximation

and in the detail coefficients. The level l=3 allowed

to retain the gamma and alpha bands of the original

signal yielding the CD

details representation in the

band (250/2

, 250/2

] Hz. After bands selection, the

trial was divided into q sub-trials having an

overlapping zone of p points (points in common

between consecutive sub-trials) to maintain

continuity between pieces. The overlapping region

has been introduced to avoid the exclusion of useful

information that could be present on the tails of

consecutive sub-trials. The classification used the

division in sub-trials in order to discard the noisy

pieces of each signal while retaining and averaging

the useful information (Petrantonakis and

Hadjileontiadis, 2011). In this set of sub-trials, a

group of

N characteristics describing the signal,

the features, was calculated; the most common

features used in literature (Subasi, 2007; Cvetkovic,

2008) and considered herein were: the mean and the

median values, f

and f

respectively, the mode f

(i.e. the most frequent value in a sub trial), the

largest and the smallest elements, f

and f

respecti-

vely, the range f

of the values and their standard

deviation f

, the mean value f

and the median f

the absolute value of the difference between the

vector and its mean value, the sum f

of all the

elements, the norm f

and the maximum value f

Since the set of the original features could be

redundant, their number was reduced to s by

applying the Principal Component Analysis (PCA)

in the modified form proposed in (Song et al., 2010).

In particular, the covariance matrix of the

standardized data was computed and its largest

eigenvalues, and its corresponding eigenvectors,

were selected. The selected features were the ones

with the higher weights in the covariance matrix.

Hence the classification of the signal was performed

through Support Vector Machine (SVM) by using

the remaining features.

Figure 2: Block diagram of the classification procedure.

The SVM determined the optimal hyperplane to

separate data in two classes, Class 1 and Class 2, and

Classiﬁcation of Emotional Signals from the DEAP dataset

it was obtained as a trade-off between the

requirement of maximizing the Euclidean distance

between the closest points and the requirement of

minimizing the error on misclassified points. The

classification method is summarized in the block

scheme of Figure 2.

The off-line step was the calibration: the most

predominant features referring to the two different

conditions, namely, to an activation stage (class #1)

or to another (class #2), were selected by conside-

ring that the class of the given signal was known in

advance; then the SVM classifier was trained. The

on-line step was the classification of a signal of

which the placement to the class #1 or #2 was

unknown in advance. More details can be found in

(Iacoviello et al., 2015c).

3 RESULT AND DISCUSSION

As specified above, the proposed classification

procedure was applied to a set of 32 trials for the E1

(NVHA) emotion, 32 for the E2 (HVHA) emotion

and 32 trials for the relaxing state R for each subject

(in the DEAP database, they corresponded to the

subjects #1, 2, 22, 24, 25, and 9 indicated by S

– S

respectively, in this study) in a single test. The

extracted signals were analyzed in order to verify the

absence of abnormal spikes. In particular, the subject

#9 (S

) was discarded since his EEG signals

contained very high spikes of difficult explanation

and reduction; a more sophisticated method to

discard a subject was proposed in (Petrantonakis and

Hadjileontiadis, 2011). For each of the selected

subjects, the same emotion was associated to at least

two videos. The corresponding signals were inserted

in the same set of trials (in the same class). Of the

whole set of trials, 16 were leaved for calibration of

the methods (in particular, 8 trials were used for

training and 8 for validation) and 48 were used to

simulate real-time classification.

Parameters for the classification method: Each

trial was divided into q=7 sub-trials, each

corresponding to 300 elements, by considering p=50

overlapping points. The number of retained sub-

trials was 5.

The SVM was applied on a subset of features

extracted by PCA on the basis of the averaged

values of the features calculated on the selected 5

sub-trials. The considered emotions were clearly

recognizable (classification accuracy close to 100%)

from the relaxing condition, for most of the

considered subjects, both from the left and from the

right hemisphere of the brain. This occurred also

when considering the classification of the two

emotions reciprocally. However, by considering just

the channels whose accuracy was exactly 100%, it

may be noted that the right hemisphere was

prevalent. Moreover, the channel PO4 recurred in all

the subjects and was useful both to classify the two

emotions from “relax” and the two emotions

reciprocally. This was particularly evident for the

subject S

that represented the worst case, among the

considered subjects, in terms of accuracy

distribution between channels (Figure 3).

Also in this worst case, the two emotions

activated mostly the right hemisphere, and the

channel PO4 was the most important (100% of

accuracy) both to recognize the considered emotions

from “relax” and to recognize the emotions each

other, though with different feature sets. In this case,

also the channel AF4 (80% in accuracy) was useful

to recognize both the emotions with respect to relax,

though it was not specific to recognize one emotion

with respect to the other. These results confirmed

what was previously found in (Placidi et al., 2015a)

and (Placidi et al., 2015b) with respect to the

activation of the right hemisphere for a negative

emotion but, differently from other works (Song et

al., 2010; Davidson et al., 1979), the brain

lateralization between the two types of emotions was

not clearly evident.

Figure 3: Accuracy values for each channel for the subject

S3 (our worst case). The reported maps are referred to

binary classifications between: negative emotion vs relax

(a), positive emotion vs relax (b), negative vs positive (c),

and negative vs flipped positive (left-right brain

hemispheres) (d), respectively.

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

For better clarifying this aspect, we also tried to

classify the signals of the first emotion with those of

the second one having flipped the two brain

hemispheres of the second emotion. In this case, we

observed that most of the channels were

recognizable to one another (accuracy close to

100%), thus demonstrating that a correlation

between channels of opposite hemispheres were

absent. In particular, Figure 3d shows the good

symmetry of the accuracy distribution for the subject

, confirming the absence of a relationship between

channels of opposite hemispheres. The activation of

a considerable number of channels could be due to

the elicitation protocol: being an external

stimulation, it would activate multiple mental

processes corresponding to the activation of

different brain regions. Conversely, in the case of

self-induced emotions (Placidi et al., 2015), the

activation, being generated by a concentration task

that reduces the possibility of distraction, was

concentrated in a specific brain location.

It is useful to evaluate the importance of the

features after the PCA selection. Table I shows the

channels in which the accuracy was above 80% and

the corresponding 4 most important features used for

classification, reported in order of descending

importance.

The most activated channels were those with

higher accuracy, expressed as the ratio between the

number of right answers with respect to the total

number of trials in percentage. It is important to note

that, due to the protocol used to collect the data of

the DEAP database (audio-videos were used to elicit

emotions), both temporal (influenced by listening

audios) and occipital (influenced by viewing videos)

channels were left out. Moreover, just accuracy

above 80% was considered.

From the analysis of Table I it can be further

observed that, even between the 4 most influent

features used for classification in different channels

there was a sort of recurrence: features 5, 7, 9 and 10

occurred very often, though with different order.

The previous results demonstrated that the

proposed method could be effectively used for

finding the proper, optimized and minimal

combination of channels/features for effective

classification of the 3 considered emotional states.

The choice of leaving separated the EEG channels in

the classification process was finalized at the

evaluation of the contribution of each channel to the

whole classification process. The obtained results

demonstrated also that the method could be applied

for effective classification of the 3 considered

emotional states by using a polling system after the

application of 3 mutual binary classification

methods.

Table 1: Best channels (accuracy above 80%) and the

corresponding 4 best features used for classification.

Temporal and occipital channels were not considered.

E1vsR E2vsR E1vsE2

Ch Features Ch Features Ch Features

AF35910 7 AF395 10 7 CP1 8 6 3 1

F3 5 9 10 2 F3 5 9 10 2 F4 9 5 1 3

C3 5 9 7 10 C3 9 5 7 10 C4 5 9 3 1

CP1 9 5 10 7 CP1 8 9 2 10 PO4 9 5 10 8

F4 5910 7 F4 45 10 3

C4 9 5 7 10 C4 9 5 7 10

PO4 5 9 7 10 PO4 5 9 7 10

CP1 5 9 7 1 FC1 5 9 10 7 FC1 9 5 3 1

FC1 9 5 3 1 CP2 5 9 10 7 C3 5 9 10 1

C3 9 5 10 7 F4 9 5 10 7 PO4 9 5 7 10

F3 5 9 7 10 PO4 9 5 7 10

CP6 9 5 7 10

C4 5 9 7 3

PO4 9 5 7 10

AF4 9 5 7 3 FC1 9 5 7 2 PO4 3 1 2 10

CP2 9 5 7 10 CP5 9 5 8 6

PO4 9 5 2 7 AF4 9 5 7 2

CP2 9 5 7 1

PO4 3 1 2 10

FC5 5 9 7 10 AF3 5 9 10 7 FC1 9 5 10 7

FC1 9 5 7 10 FC1 9 5 7 10 CP1 5 9 10 7

FC2 5 9 7 10 CP5 4 5 10 7 PO3 9 5 10 7

P4 5 9 7 10 CP6 5 9 2 10 CZ 9 5 3 1

PO4 5 9 7 4 PO4 9 5 7 4 FC6 6 8 10 7

C4 5 9 1 3

PO4 5 9 10 7

AF39510 7 AF395 10 7 F3 9 5 7 3

F3 5 9 7 10 C3 9 5 7 10 FC5 9 5 7 10

AF4 5 9 3 1 CP1 5 9 10 7 CP1 5 9 10 7

C4 5 9 10 7 AF4 5 9 7 10 FC6 9 5 10 7

PO45910 7 F4 95 10 7 C4 5 9 7 10

CP2 5 9 10 7 PO4 5 9 10 7

PO4 5 9 10 7

The choice of considering simple binary

classifications instead of more articulated multi-class

strategies (ternary in our case) was twofold: first,

binary classification was really simple and its results

could be more reliable than those of a multi-class

strategy (in that phase, we were still trying to

comprehend the neurophysiologic complex mecha-

nisms and the brain/features mapping of different

emotions and we would avoid biases from the

classification strategy); second, the extension of the

proposed method to the classification of more than 3

classes could be very simple (the addition of a class

simply involves the addition of a series of binary

classifiers). Finally, the results showed that the

proposed method could be easily applied for

Classiﬁcation of Emotional Signals from the DEAP dataset

classification of emotions that have similar spatial

activation pattern because they could be recognized

through their specific features combination.

Regarding the computational time, the algorithm

was implemented in Matlab® on a personal

computer (Intel(R) Core(TM) i7-4790 CPU @3.60

GHz 3.60 GHz RAM: 16,0 GB): the classification

method took 17.2 minutes for processing all the

trials used for the calibration step, 36.4 minutes for

the training, 35 sec for the validation and 2.2x10-

sec for processing a single trial. Time could be

strongly reduced by using just a single channel for

classification (PO4).

4 CONCLUSION AND FUTURE

DEVELOPMENTS

A comparison of EEG emotional signals from differ-

rent subjects of the DEAP database was performed

by using a competitive machine learning based

classification method. The compared emotions were

NVHA and HVHA both with the relaxing state and

reciprocally.

The obtained results showed that, though

defining different feature sets for different channels,

most of the measured channels allowed high

classification accuracy. This was true both by

comparing each emotion with “relax” and by

comparing the two emotions. In order to highlight if

there was a lateralization between the brain

hemispheres when subjected to opposite emotions,

we also compared the data of the first emotion with

those of the second after the flipping of the right

hemisphere data with those of the left hemisphere of

the second emotion. The results showed that, being

the classification accuracy very high for most of the

channels, the activation was not differently

distributed between hemispheres for different

emotions. The obtained results demonstrated that the

proposed classification method could be efficiently

applied both to discover the neurophysiological

mechanisms of different emotional states and to

efficiently recognize the minimal channels/features

set for recognizing each of the considered emotions

from the others and from the relaxing state.

Moreover, the method could be efficiently used for

classification of an incremental number of emotions

through the further introduction of a set of binary

classifiers and a proper polling scheme.

Future work will be spent in

1) refining the strategy of subjects selection

(Petrantonakis and Hadjileontiadis, 2011),

2) mapping the channels/features pattern of

new emotional states,

3) inserting the previous emotional states in

the classification process,

4) classifying the considered emotional states

in groups (more than two emotions at once)

5) exploring a deeper wavelet prefiltering,

assuming the decomposition that better

enhance the significant part of the elicited

signal,

6) selecting the most informative subset of

channels,

7) selecting a strategy of self-induction of a

set of emotions in order to allow also the

usage of the neglected channels,

applying the obtained channels-features maps to

implement a classification strategy for multi-

emotions based BCIs to be used for communication

purposes or for affective computing applications.

ACKNOWLEDGEMENTS

The authors wish to thank Mrs Carmelita Marinelli

for her help in preparing the figures.

REFERENCES

Shih, J., Krusienski, D. and Wolpaw, J., 2012. Brain-

Computer Interfaces in Medicine. Mayo Clinic

Proceedings, 87(3), pp.268-279.

Kubler, A., Nijboer, F., Mellinger, J., Vaughan, T.,

Pawelzik, H., Schalk, G., McFarland, D., Birbaumer,

N. and Wolpaw, J., 2005. Patients with ALS can use

sensorimotor rhythms to operate a brain-computer

interface. Neurology, 64(10), pp.1775-1777.

Hochberg, L., Serruya, M., Friehs, G., Mukand, J., Saleh,

M., Caplan, A., Branner, A., Chen, D., Penn, R. and

Donoghue, J., 2006. Neuronal ensemble control of

prosthetic devices by a human with tetraplegia.

Nature, 442(7099), pp.164-171.

Fisch, B. and Spehlmann, R., 1999. Fisch and

Spehlmann's EEG primer. Amsterdam: Elsevier.

Babiloni, F., Cincotti, F., Lazzarini, L., Millan, J.,

Mourino, J., Varsta, M., Heikkonen, J., Bianchi, L.

and Marciani, M., 2000. Linear classification of low-

resolution EEG patterns produced by imagined hand

movements. IEEE Transactions on Rehabilitation

Engineering, 8(2), pp.186-188.

Placidi, G., Avola, D., Petracca, A., Sgallari, F. and

Spezialetti, M., 2015a. Basis for the implementation of

an EEG-based single-trial binary brain computer

interface through the disgust produced by

remembering unpleasant odors. Neurocomputing, 160,

pp.308-318.

NEUROTECHNIX 2016 - 4th International Congress on Neurotechnology, Electronics and Informatics

Pistoia, F., Carolei, A., Iacoviello, D., Petracca, A., Sacco,

S., Sarà, M., Spezialetti, M. and Placidi, G., 2015.

EEG-detected olfactory imagery to reveal covert

consciousness in minimally conscious state. Brain

Injury, 29(13-14), pp.1729-1735.

Placidi, G., Petracca, A., Spezialetti, M. and Iacoviello,

D., 2015b. Classification strategies for a single-trial

binary Brain Computer Interface based on

remembering unpleasant odors. 2015 37th Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC).

Iacoviello, D., Petracca, A., Spezialetti, M. and Placidi,

G., 2015a. A real-time classification algorithm for

EEG-based BCI driven by self-induced emotions.

Computer Methods and Programs in Biomedicine,

122(3), pp.293-303.

Iacoviello, D., Pagnani, N., Petracca, A., Spezialetti, M.

and Placidi, G., 2015b. A Poll Oriented Classifier for

Affective Brain Computer Interfaces. Proceedings of

the 3rd International Congress on Neurotechnology,

Electronics and Informatics.

Iacoviello, D., Petracca, A., Spezialetti, M. and Placidi,

G., 2015c. A Classification Algorithm for

Electroencephalography Signals by Self-Induced

Emotional Stimuli. IEEE Trans. Cybern., pp.1-10.

Placidi, G., Petracca, A., Spezialetti, M. and Iacoviello,

D., 2015c. A Modular Framework for EEG Web

Based Binary Brain Computer Interfaces to Recover

Communication Abilities in Impaired People.J Med

Syst, 40(1).

Guler, I. and Ubeyli, E., 2007. Multiclass Support Vector

Machines for EEG-Signals Classification.IEEE

Transactions on Information Technology in

Biomedicine, 11(2), pp.117-126.

Koelstra, S., Muhl, C., Soleymani, M., Jong-Seok Lee,

Yazdani, A., Ebrahimi, T., Pun, T., Nijholt, A. and

Patras, I., 2012. DEAP: A Database for Emotion

Analysis; Using Physiological Signals. IEEE Trans.

Affective Comput., 3(1), pp.18-31.

Russell, J., 1980. A circumplex model of affect. Journal of

Personality and Social Psychology, 39(6), pp.1161-

1178.

Daubechies, I., 1992. Ten lectures on wavelets.

Philadelphia, Pa.: Society for Industrial and Applied

Mathematics.

Mamun, M., Al-Kadi, M. and Marufuzzaman, M., 2013.

Effectiveness of Wavelet Denoising on

Electroencephalogram Signals. Journal of Applied

Research and Technology, 11(1), pp.156-160.

Subasi, A., 2007. EEG signal classification using wavelet

feature extraction and a mixture of expert model.

Expert Systems with Applications, 32(4), pp.1084-

1093.

Cvetkovic, D., Übeyli, E. and Cosic, I., 2008. Wavelet

transform feature extraction from human PPG, ECG,

and EEG signal responses to ELF PEMF exposures: A

pilot study. Digital Signal Processing, 18(5), pp.861-

874.

Song, F., Guo, Z. and Mei, D., 2010. Feature Selection

Using Principal Component Analysis. 2010

International Conference on System Science,

Engineering Design and Manufacturing

Informatization.

Petrantonakis, P. and Hadjileontiadis, L., 2011. A Novel

Emotion Elicitation Index Using Frontal Brain

Asymmetry for Enhanced EEG-Based Emotion

Recognition. IEEE Transactions on Information

Technology in Biomedicine, 15(5), pp.737-746.

Davidson, R., Schwartz, G., Saron, C., Bennett, J. and

Goleman, D., 1979. Frontal versus parietal EEG

asymmetry during positive and negative affect.

Psychophysiology, 16(2), pp.202-203.

Classiﬁcation of Emotional Signals from the DEAP dataset