Galvanic Skin Response Data Classification for Emotion Detection

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 8, No. 5, October 2018, pp. 4004~4014
ISSN: 2088-8708, DOI: 10.11591/ijece.v8i5.pp4004-4014  4004
Journal homepage: http://iaescore.com/journals/index.php/IJECE
Galvanic Skin Response Data Classification for
Emotion Detection
Djoko Budiyanto Setyohadi1
, Sri Kusrohmaniah2
, Sebastian Bagya Gunawan3
, Pranowo4
,
Anton Satria Prabuwono5
1,3,4
Departement of Informatics Engineering. Universitas Atma Jaya Yogyakarta, Yogyakarta 55281, Indonesia
2
Psychology Department, Universitas Gajah Mada, Yogyakarta, Indonesia
5
Faculty of Computing and Information Technology, King Abdulaziz University, Rabigh 21911, Saudi Arabia
Article Info ABSTRACT
Article history:
Received Feb 12, 2018
Revised May 25, 2018
Accepted Jul 2, 2018
Emotion detection is a very exhausting job and needs a complicated process;
moreover, these processes also require the proper data training and
appropriate algorithm. The process involves the experimental research in
psychological experiment and classification methods. This paper describes a
method on detection emotion using Galvanic Skin Response (GSR) data. We
used the Positive and Negative Affect Schedule (PANAS) method to get a
good data training. Furthermore, Support Vector Machine and a correct
preprocessing are performed to classify the GSR data. To validate the
proposed approach, Receiver Operating Characteristic (ROC) curve, and
accuracy measurement are used. Our method shows that the accuracy is
about 75.65% while ROC is about 0.8019. It means that the emotion
detection can be done satisfactorily and well performed.
Keyword:
Classification
Emotion detection
Experimental research
Galvanic skin response
Support vector machine Copyright © 2018 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Djoko Budiyanto Setyohadi,
Magister Teknik Informatika,
Universitas Atma Jaya Yogyakarta,
Jl. Babarsari 43 Yogyakarta 55281, Indonesia.
Email: djoko.bdy@gmail.com
1. INTRODUCTION
Currently, human emotion is an important thing and has become a concern in various fields of
research, including computer science [1]. This phenomenon is real, considering that emotion as a human
expression implies activities done by human beings and their excitement can be detected based on the data
measured from them: voice, face, body, and physiology [2], [3]. Detection of human emotion is potential to
render the development of diverse aspects of life such as entertainment, learning, and other elements [2],
[4]-[6]. The detection is part of a data mining process, namely classification, and clustering. Until today,
there have been at least ten machine learning algorithms which can be used for both classification and
clustering [7]. The following methods are examples of the detection algorithm based on classification and
clustering, which means classifying like Support Vector Machine (SVM), Naïve Bayes, K-Nearest Neighbors
(KNN), Decision Tree and many others. They reported that the performance of classification could be
enhanced when they customize the parameter within the algorithm to gain a maximum result [7].
SVM is one of the popular algorithms due to its sufficient capability in nonlinear classification,
regression and outlier detection even before the algorithm became popular [7]. Although the use of SVM is
more straightforward than other algorithms, to obtain an optimal result, an appropriate approach is required.
Some methods that have been commonly applied include transformation, feature scaling, choosing the right
kernel and finding the right C and γ [8]. SVM works by detecting the hyperplane that has the most significant
margin when a dimension is more than two (2), so data can be effectively separated in a linear manner (this
projection is carried out in kernel technique). There are four primary kernels in SVM, namely linear, radial

Int J Elec & Comp Eng ISSN: 2088-8708 
Galvanic Skin Response Data Classification for Emotion Detection (Djoko Budiyanto Setyohadi)
4005
basis function (RBF), sigmoid, and poly-nominal. Among the four kernels, the RBF is worth considering to
be used due to its capability in separating nonlinear data [9], [10]. With the capacity, the use of SVM can be
employed as a promising alternative for emotion detection.
Emotion detection using biosensor has become popular recently. Besides its frequent use in many
applications, bio censor also possesses the best performance indirectly measuring parameters in the human
body, whose value is influenced by human emotion status. These sensors can measure generated-signals from
human body such as heart rate, blood pressure, temperature, skin conductance, etc. Many scholars reported
that the variety of measurement of generated-signals could be used as a foundation of emotion detection.
[11]-[14], and one of the favorite bio censor nowadays is Galvanic Skin Response (GSR).
Galvanic Skin Response (GSR) is a tool which censors human psychological symptom. It measures
signals sent by the human skin which is considered as a reflection of physiological changes [15], [16]. The
use of GSR for physiological purposes is firstly implemented by Vigouroux. Moreover, many researchers
reported the use of GSR for physiological phenomena [15], [17]. GSR signal indicates electrical changes
measured at the surface of the human skin that varies with the changes in skin moisture level (sweating) and
reflects the differences in their sympathetic nervous system [15]. The GSR signals are formed by the
changing body resistance due to the changing body conductivity caused by the production of sweat. This tool
has been quite favorite as it is adequately sensitive and inexpensive. In its development, the use of GSR to
detect emotion can be carried out due to the spontaneous reaction which cannot be controlled by the user.
Furthermore, it is considered as one of the strongest signals that can be used for emotion detection [18], [19].
Relying on the classification performance of SVM in the previous research particularly when SVM should
deal with the temporal data classification, SVM was taken as the SVM algorithm to detect emotion based on
GSR data.
Within this decade, a lot of researchers on emotion detection have been conducted. The algorithm of
various detection tools such as GSR, electroencephalogram (EEG) and other devices have been made [3], [4],
[12]. For example, EEG is used for measuring the change of positive and negative emotion of the subject
who is induced through an IDFA or qDFA video. It was reported that the accuracy reached 74.3% while the
accuracy of the qDFA model reached 69.4%. It was also said that if it was applied to detect personal emotion,
the accuracy could reach 94.5% [20]. The combination of EEG with human physiological change called Eye
Movement Signals was also done [21]. They reported that when the combinations of the two factors are
paired fuzzy integral fusion algorithm classification could increase the performance up to 10% so that the
system accuracy became 87.59.EEG data is interesting to process and tend to have high accuracy in the
context of emotion detection problems. However, EEG has several crucial issues, which are (1) expensive
and (2) high noise of data resulted by its sensor. This phenomenon has made researchers attempt to find
alternatives that are more appropriate, and GSR is one of the other options to choose. GSR is an alternative
that has been developed since it is cheap and adequately sensitive. The two reasons at least were related to
the main reason why GSR is applied to detect psychological condition using various methods. There are
several studies which propose different methods of detecting stress levels by measuring skin conductance
[14], [22]. An example of this research was the one conducted by Villarejo. They attempted to classify stress
suffered by a person using GSR by providing stress induction in a mathematical calculation. The result was a
model that was formed by using GSR data could classify stress with an accuracy that reached up to
76.56% [22].
The study described in [23] has the objective of detecting sweat levels for the diagnosis of
sudomotor dysfunction, something that can help in the diagnosis of diabetes. There are other medical
applications based on skin conductance, such as epilepsy control: sweaty hands may be a warning signal of
an epileptic attack, as support of the diagnosis and treatment of bipolar disorder patients [24]. By combining
the sweat of the hands with the temperature of the skin, it is possible to develop a truth meter [18]; as when
the person is lying, his hands are colder, and skin resistance is lower. In this case, it is not necessary to
include an ADC because the variation of skin resistance happens at odd times so, with different resistances
and transistors, it is possible to build a lie detector. A truth meter is possibly established based on the hands
sweat and skin temperature. Person's hands are colder, and he has lower skin resistance when he is telling a
lie.
Attempting to do the same thing done by Villarejo, Kurniawan classified stress using GSR and noise
signals [14], [22]. Classification by noise signals generated the best result, which was 92% through SVM
algorithm. When the GSR data was used, however, the accuracy obtained was only 70% through the same
algorithm. There was an opinion that when we attempted to classify a person's stress using another person's
data, the generated accuracy would decrease [14]. Researchers commonly use GSR to classify the human
stress level. In a different approach, Liu attempted to classify emotional status into happiness, grief, fear,
anger, and calm using GSR data. The best accuracy level obtained by Liu was 66.67% through SVM
algorithm with kernel radial based function (RBF) [12]. Recently, the Sliding Window technique has often

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 8, No. 5, October 2018 : 4004 – 4014
4006
been used in physiological data classification. In the research, the stress can be appropriately classified,
where the accuracy of classification is up to 98% by combining HRV data with ECG data which were
processed by Sliding Windows using widths 24 points, and it was classified using SVM with linear kernel
[25]. A different method was carried out by Wei, who attempted to classify emotion into anger, fear, joy,
sorrow, acceptance, rejection, surprise and expectancy. The research obtained a result with reasonable
accuracy as much as 80% using SVM algorithm which was processed by Sliding Windows technique [12].
Using different approach, Guo also tried to classify human emotion into amusement, fear, relaxation, and
sadness using Sliding Windows technique with the lag as much as 20 points. The research also obtained a
good result with accuracy that reached 79.45% [18].
Support Vector Machine (SVM) was first introduced by Vladimir Vapnikin 1995 and used for
binary classification purpose [26]. When SVM performs, it will create hyperplane which is aimed as a divider
of each class. The basic concept was looking for the hyperplane which separated d-dimensional data correctly
into two classes. The hyperplane had the biggest margin. For further explanation see Figure 1.
(a) (b)
Figure 1. Support vector machine (a) and an example of kernel solving (b)
The following formula is to calculate the hyperplane in the SVM:
  iii bwxy  1 (1)
in which yiis label, xiis input, w is weight and b is bias.
Although the SVM has proved to become a highly effective learning algorithm in the real world
data, it has not frequently been used. It is caused by several problems found in the SVM, such as the fact that
the implementation of the SVM which is not easy. Besides, the use of the wrong parameter may increase
difficulty level in the computation process. Initially, the SVM was created with a method of linear problem-
solving. However, there have been data that are not linear. Therefore, kernels that could be able to map the
data in the dimensional space which is usually bigger (feature space) were created. Accordingly, nonlinear
problems that could previously not be adequately solved could eventually be solved. There have been kernels
that frequently used in SVM such as radial basis function (RBF) or Gaussian, polynomial and sigmoid. The
figure shows two an example of problem-solving using kernels. However, the use of kernels has particular
weaknesses. Wrong measuring parameters in kernels can cause the model to become over lift or underfit.
Also, the use of kernels can increment the computation load so that it causes the process to slow down. RBF
is one of the kernels, which has frequently been used since it has a good performance and tested. RBF maps
nonlinear data into a bigger dimension. The hyperplane in the nonlinear data can be sought. The formula of
RBF kernels data mapping is as follows:





 
 2
2
2
||||
exp),(

xbxa
xxK ba
(2)
in which  can be changed according to the needs [27].

4007
2. PROPOSED METHODOLOGY
The research methodology of the data collection was divided into two parts. One was designed for
the respondents, and the other was intended for the system. The design for the respondents used the with-
subject designs, and the system design was developed by combining several methodologies that corresponded
to emotional detection using audio-visual induction. In this study, as many as 39 respondents were chosen
randomly from the Psychology Faculty of GadjahMada University, Yogyakarta and the Industrial
Technology Faculty of Atma Jaya University, Yogyakarta, all of whom were still active students.
Methodology, as an overall process, is illustrated in Figure 2. While the data collection process was done in
an isolation room and the tool installation is configured in Figure 3.
PANAS scale measurement
Data Collection
Preprocessing
Testing and validation
Parameter Selection and
Model Training
Figure 2. Research methodology Figure 3. The GSR installation
2.1. PANAS scale measurement
PANAS SCALE was given at the beginning of the experiment to obtain both positive and negative
affection baselines. The baseline of physiological sensor data was recorded for three minutes after the subject
had finished working on the PANAS scale. During the three minutes, the respondents were required to wait
for the next instruction. After three minutes, an instruction would appear on the subject to memorize
unrelated pairs of words followed by the actual task given to them, which was to remember the pairs of
words that had been memorized using the cued recall method. PANAS Scale is a self-report questionnaire
consisting of 20 items with answers choices in the Likert scale form of 5 answer choices used to measure the
level of affection [28]. We had prepared a schedule of data collection which was conducted for six days,
starting from 08.30 to 16:30 with 30 minutes between sessions. Subject grouping in the experimental and
control groups was carried out randomly, and so were the presentation of stimuli and lists of words to
remember. Before data collection, each subject was required to fill in the informed consent containing the
research procedures to be followed, the consequences that might occur, as well as the incentives that the
subjects would receive. If they agreed, they were asked to sign the informed consent. After that, the
experimenters gave an additional explanation to ensure that the subjects completely understood the
experiment procedure.
2.2. Data collection
Data collection was done in a room which complied with the experimental protocol. The settings
here included the temperature, light and noise set. The treatment of the condition and equipment was the
same for all respondents, which included room, computer, screen, speakers, GSR sensor, keyboard, mouse,
table, and chair.
2.3. Pre-processing
Pre-processing was divided into several stages, namely data categorization, data aggregation, data
normalization, and data lagging. The data obtained were then categorized into three columns by the treatment
given at the time of the experiment (positive, neutral, negative). This data categorization would be useful in
data labeling. The next process was to aggregate the data to convert the time unit into seconds by averaging
the data according to the GSR frequency at the time of data collection. In this case, the data was averaged on
every 103 data. The next thing to do was normalization or feature scaling which was useful for standardizing
the distance or range of variables or features in the data so that the maximum and minimum value became 1
and 0. Finally, the data was formed by the number of n lag and was added with a positive, neutral or
negative label.

 ISSN: 2088-8708
4008
)min()max(
)min(
xx
xx
x



(3)
2.4. Parameter selection and modeling training
In this research, the model was trained to use SVM algorithm. The number of lags that was used
started from 5 to 40 lags SVM parameter tuned automatically with the cost starting from 2^1 to 2^8 using the
linear kernel, polynomial, radial based function and sigmoid.
2.5. Model testing and validation
After the parameter required by SVM model was set, the result was validated to approve the quality
of the classification model. The accuracy and Receiver Operating Curve were obtained by using the
following formula:
FNFPTNTP
TNTP
Accuracy


/
(4)
where: TP is true positive, TN is true negative, and FP is false positive and FN is false negative
3. RESULTS AND ANALYSIS
In this section, it is explained the results of research and at the same time is given the
comprehensive discussion. Results can be presented in figures, graphs, tables and others that make the reader
understand easily [2], [5], [29]. The discussion can be made in several sub-chapters.
3.1. Data set validation based on PANAS scale
Manipulation validation was done by running Repeated Measures ANOVA on the PANAS scale.
Manipulation can be declared successful if there is a different score of PANAS in a condition that is relative
to the baseline and also a difference in each condition of treatment. The descriptive data of the two sub-scales
of PANAS can be seen in Table 1. Sphericity validation of positive and negative affection sub scale of
PANAS scale both suggested coefficient p<0.001. Correction on sphericity was done by employing
Greenhouse-Geisser (since the epsilon was above 0.7).
Table 1. Statistical Description of PANAS Scale of Positive and Negative Affections
Positive Affection Negative Affection
Mean SD Min Max Mean SD Min Max
Baseline 3.566 0.4828 2.3 4.3 2.171 0.6182 1.2 4
Neutral Condition 3.229 0.6049 1.5 4.4 2.021 0.748 1 4.1
Positive Condition 3.284 0.5659 1.8 4.2 1.992 0.7585 1 4.4
Negative Condition 3.239 0.6262 1.4 4,2 2.339 0.7387 1.1 4.1
It is seen in Table 2 that the result of Repeated measures ANOVA with Greenhouse-Geisser
correction resulted from coefficient p<0.001 in the sub-scale of positive affection and p=0.002 (p<0.01) in
the sub-scale of negative affect, which meant that there were significant differences between the two in at
least one of the conditions. Validation that is more detailed was done by carrying out post hoc using
Bonferroni method, of which result can be seen in Table 3.
Table 2. Result of Repeated Measures ANOVA with Sphericity Correction; **p <0,001
Sub-scale Sphericity Correction ε df p
Positive Affection Greenhouse-Geisser 0.705 2.114 <0.001**
Negative Affection Greenhouse-Geisser 0.763 2.290 0.002**
Table 3. Post Hoc Bonferroni Validation of PANAS Scale; *p <0.05; **p <0.01
Mean difference pbonf Mean Difference pbonf
Baseline Neutral 0.337 0.011* 0.150 0.688
Positive Condition 0.282 0.037* 0.179 0.332

4009
Mean difference pbonf Mean Difference pbonf
Negative Condition 0.326 0.011* -0.168 0.848
Neutral Condition Positive Condition -0.055 1.000 0.029 1.000
Negative Condition -0.011 1.000 -0.318 0.002**
Positive condition Negative Condition 0.045 1.000 -0.347 0.004**
In the PANAS sub-scale of positive affection, significant difference (p < 0.05) occurred in the
baseline condition rather than in neutral condition (p = 0.011), positive condition (p = 0.037), and negative
condition. It meant that providing the treatment had successfully changed the condition of the subject from
the baseline condition based on the result of the PANAS scale. Descriptively, the condition of positive
affection treatment showed higher mean (3.284) than that in the neutral condition (mean = 3.229) and
negative condition (mean = 3.239). The difference was not statistically significant (p > 0.05), which meant
that giving the treatment was inadequately effective in inducing the valence of different kinds of affection.On
the PANAS sub-scale of negative affection, significant differences occurred in negative conditions rather
than in positive condition (p = 0.004; p <0.01) and neutral condition (p = 0.002; p <0.01). It indicated that
based on the PANAS scale, the provision of negative affection treatment had successfully induced negative
affection, relative to the other two conditions. Negative treatment condition descriptively increased negative
affection rather than the baseline condition (mean difference = -0.222) although it was not statistically
significant. The researchers concluded that the provision of affection induction had successfully changed the
positive affection condition from the baseline although it was not yet affective to differentiate it from other
treatment condition. Meanwhile, the result of the negative affection sub-scale suggested that negative
treatment condition could induce negative affection relative to other treatment conditions. There were
possible measurement errors caused by the positive affection sub-scale of PANAS scale since some items
had a relatively low inter-correlation among them. Furthermore, the process of classifying the GSR data to
see the emotional status of the respondents will be discussed in the next chapter.
3.2. Classification of GRS data
The data resulted by the GSR were categorized into three, namely positive, neutral and negative
according to the condition in time of data collection. The result of the data category was still in the form of
time series. For the sake of SVM algorithm training, the data would be separated by class or cut by the
category. Examples of the data categorization can be seen in Figure 4, Figure 5 and Figure 6.
Time (Frequency)
Skin resistance
Figure 4. Example of positive data
Skin resistance
Time (Frequency)
Figure 5. Example of neutral data
Time (Frequency)
Skin resistance
Figure 6. Example of negative data

 ISSN: 2088-8708
4010
3.3. Preprocessing
There are three data preprocessing performed in this research, i.e., Aggregation,
Normalization/Feature Scaling, Lagging, and Labelling. Data aggregation was done by changing the time
unit from frequency into second. In data mining, aggregate data is useful and is aimed to increase the
signal/noise ratio commonly, especially on the temporal data processing. The result of the data aggregation
can be seen in Figure 7. The result showed that the data still have the same patterns with those before the
aggregation. Normalization is a rescaling of the data from the original range so that all values are within the
range of 0 and 1.
Normalization can be useful and even required in some machine learning algorithms when your time
series data has input values with differing scales. It may be needed for any classification algorithms which
use distance as a foundation of data processing. The result of data normalization in the research can be seen
in Figure 8. The last preprocessing is Lagging and Labelling. Lagging on data functions is preparing and
framing the time series data. Example of lagging and labeling on data with five lags can be seen in Table 4.
Table 4. Example 5 rows of Data Lagging and Labeling
lag.1 lag.2 lag.3 lag.4 lag.5 label
0.109827 0.216149 0.334738 0.539203 0.613395 positive
0.216149 0.334738 0.539203 0.613395 0.929314 positive
0.334738 0.539203 0.613395 0.929314 1 positive
0.539203 0.613395 0.929314 1 0.812889 positive
0.613395 0.929314 1 0.812889 0.717254 positive
3.4. Model and validation
In this research, the processed data were classified with different kinds of lags and kernels to see the
comparison of accuracy rate between the kernels and different numbers of lags. The data were classified by
using a library "e1071" in language R. For the training, random sampling as much as 70% was used and the
validation used the rest 30% of the data. The classification started with the 5 to 41 lags. The result of
accuracy comparison in the classification model can be seen in Figure 9.
Negative
Positive
Neutral
Figure 9. ROC graph
Time (seconds)
Skin resistance
Figure 7. Example of positive aggregation
Time (seconds)
Skin resistance
Figure 8. Example of normalization of positive data

4011
It can be seen in Figure 9 that the accuracy level tends to increase along with the increasing number
of lags. It is also noticed that the RBF kernels suggested the best performance compared to the kernels and
algorithms of other classifications. A detailed result of this process is shown in Table 5. This table shows that
SVM model using RBF kernels is the best, which was 75.65%. The graph of ROC curve was visualized the
classification result with the highest accuracy, which was 75.65%, and it can be seen in Figure 9. Figure 9
shows the mean of the AUC is 0.8019. It indicated that the result of the emotion classification was quite good
as the mean of the AUC was close to 1.
Table 5. The Percentage Accuracy Comparison
Lag RBF Linear Naïve Bayes KNN
24 0.5220 0.4389 0.4627 0.4915
25 0.5589 0.4285 0.4660 0.4875
26 0.5932 0.4500 0.4858 0.5122
27 0.6068 0.4371 0.4730 0.5329
28 0.6157 0.4161 0.4288 0.5265
29 0.5896 0.4285 0.4421 0.5328
30 0.6262 0.4101 0.4927 0.5072
31 0.6073 0.3795 0.4790 0.5418
32 0.6080 0.4005 0.4375 0.5056
33 0.6223 0.4055 0.4953 0.5294
34 0.5939 0.4095 0.4368 0.5802
35 0.7110 0.3878 0.4410 0.4790
36 0.6538 0.3803 0.4529 0.5213
37 0.6520 0.4362 0.4852 0.5343
38 0.6322 0.2816 0.4310 0.4712
39 0.7083 0.4583 0.4930 0.5069
40 0.7565 0.4260 0.5478 0.5391
41 0.7529 0.3411 0.4941 0.5764
4. DISCUSSION
From the results obtained, it appeared that the higher the lag was, the higher the accuracy would
tend to be in the RBF kernel and it was linear. It was known in the RBF kernel that to perform the
classification of human emotional status, a large GSR pattern was required, which was 40 lags or 40 seconds
(1 lag consists of 1-second data) so that the emotional status could be well classified. Similar results were
also encountered by Grimes and Li in their previous studies. They obtained accuracy that tended to increase
along with the increasing number of the lags that were used. However, Li experienced a decrease in accuracy
and argued that the use of large lag would cause the computing process to become larger and not all datasets
would benefit as the number of lags used [30], [31]. In his research, Widodo argued that determining the
number of lag was an essential issue since too large lags could carry irrelevant information while too small
lags might discard relevant information [32], [33]. It was also mentioned by Liu in her/his research. She/He
said that the use of too large lags could disrupt the classification process because it made the data irrelevant
[25]. Widodo's and Liu's opinion can be used to explain the incidence experienced by Li, where accuracy
would increase along with increasing the number of lags but would decrease after passing a certain number
of lags. This condition is different from the opinion given by Grimes and Solovey. They argued that lag was
notable when the classification would be operated in real-time [30], [34]. Thus, the lag required to form the
relevant data in classifying human emotional status was 40 since it could generate the highest accuracy in this
classification, which was 75.65%.
Normalization using feature scaling method also directly affected the results of classification.
Classification without going through the feature scaling process could only achieve accuracy as much as
60.87%, but with the use of scaling feature the accuracy could reach up to 15% increase, amounting to
75.65% because the process of scaling feature made the data on each respondent equivalent. Accordingly, the
classification process was not disturbed by the data range that was too large or too small. This opinion was
supported by Hsu who stated that performing feature scaling before classification using the SVM was
essential as data with too large range could dominate data with a smaller range. Also, feature scaling process
also facilitated the computation process [10]. Berg specifically conducted a study on comparing the impacts
of various kinds of pre-processing on biological data. From the research, it could be concluded that the pre-
processing had big implications on biological data, especially auto scaling and feature scaling. Prevoius result
also said that the feature scaling process was related to biological data [35], [36]. Comparison of the impacts
of normalization on the results of the classification of human emotional status can be seen in Table 6.

 ISSN: 2088-8708
4012
Table 6. Effect of Normalization to an Accuracy
Normalization Accuracy
With 75.65
Without 60.87
In addition to data processing, the use of the kernel on SVM also generated different results. In
comparison, the RBF kernel could provide classification results with an accuracy of 75.65%, while the results
of accuracy by using other kernels such as linear, polynomial and sigmoid was under the accuracy level
which the RBF kernel generated. Comparison of kernel usage to accuracy result of the classification can be
seen in Table 7. Classification result by using RBF was also validated by using k Fold Cross Validation by
using k as much as 10. After doing ten times of validation using the method, the accuracy value obtained was
73% at the minimum and 78% at the maximum. The result showed that the data used in the classification
process was not linear and the RBF kernel was able to generate the best accuracy in the classification of time
series data using sliding windows technique. This result is in line with the opinion stated by Hsu, which said
that the RBF kernel had often been used to classify nonlinear data in a bigger dimensional space [10].
Table 7. Comparison of Data Accuracy to different Kernels
Kernel Accuracy
RBF 75.65
Polynomial 53.04
Linear 41.74
Sigmoid 28.7
5. CONCLUSION
We have developed a model of a system for detecting emotion from galvanic skin respond data. Our
research aimed to develop a model which can use sensor information to detect the emotion status of the
respondents. PANAS scale is applied to validate the data set for the training process. Once verified, a dataset
is used for training SVM. The experiments on emotion determination from audio-visual induction to
emotional detection showed the performance of the system. The result indicates that the model is well
performed since the accuracy is about 75.65% after some pre-processing is applied. More exploration is also
required to determine the lag of the data due to the variability in the respondents.
ACKNOWLEDGMENTS
I would like to thank Universitas Atma Jaya Yogyakarta, Indonesia for the financial support to our
research project. I would also like to express my gratitude to all members of Cognitive and Affective
Neuroscience Research Group, Psychology Department, Gadjah Mada University for providing time to
conduct an experimental process of our research.
REFERENCES
[1] N. Sebe, et al., “Multimodal Emotion Recognition”, Handbook of Pattern Recognition and Computer Vision,
vol. 4, pp. 387-419, 2005.
[2] B. Meuleman and K. R. Scherer, “Onlinear appraisal modeling: An application of machine learning to the study of
emotion production”, IEEE Transactions on Affective Computing, vol. 4, no. 4, pp. 398-411, 2013.
[3] E. H. Jang, et al., “Emotion Classification based on Bio-signals Emotion Recognition using Machine Learning
Algorithms”, vol. 3, pp. 1373-1376, 2014.
[4] K. Ansari-Asl, et al., “Channel Selection Method for Eeg Classification in Emotion Assessment based on
Synchronization Likelihood”, European Signal Processing Conference (EUSIPCO), 2007.
[5] Y. Liu, et al., “Real-time EEG-based Human Emotion Recognition and Visualization”, CW '10 Proceedings of the
2010 International Conference on Cyberworlds, pp. 262-269, 2010.
[6] F. Ringeval, et al., “Prediction of Asynchronous Dimensional Emotion Ratings from Audiovisual and Physiological
Data”, Pattern Recognition Letters, vol. 66, pp. 22-30, 2015.
[7] X. Wu, et al., “Top 10 algorithms in data mining”, Knowledge and information systems, vol. 14, no. 1, pp. 1-37,
2008.
[8] H. Y. Huang and C. J. Lin, “Linear and Kernel Classification: When to use which?”, Proceedings of the 2016 SIAM
International Conference on Data Mining, Society for Industrial and Applied Mathematic, pp. 216-224, 2016.
[9] C. Cortes and V. Vapnik, “Support-Vector Networks, Machine Learning”, Kluwer Academic Publishers, Boston,
pp. 273-297, 1995.

4013
[10] C. W. Hsu, et al., “A Practical Guide to Support Vector Classification”, 2016.
[11] Shimmer, “Measuring Emotion: Reactions to Media”, Dublin, Ireland, 2015.
[12] M. Liu, et al., “Human Emotion Recognition based on Galvanic Skin Response Signal Feature Selection and
SVM”, Smart City and Systems Engineering (ICSCSE), International Conference on, pp. 157-160, 2016.
[13] K. H. Kim, et al.., “Emotion Recognition System using short-term Monitoring of Physiological Signals”, Medical
and Biological Engineering and Computing, vol. 42, no. 3, pp. 419-427, 2004.
[14] H. Kurniawan, et al., “Stress Detection from Speech and Galvanic skin response Signals”, Computer-Based
Medical Systems (CBMS), IEEE 26th International Symposium on, pp. 209-214, 2013.
[15] N. Nourbakhsh, et al., “Using Galvanic Skin Response for Cognitive Load Measurement in Arithmetic and
Reading Tasks”, 24th Australian Computer Human Interaction Conference, pp. 420-423, 2012.
[16] P. Shangguan, et al., “The Emotion Recognition Based on GSR Signal by Curve Fitting”, J. Information and
Computational Science, vol. 11, no. 8, pp. 2635-2646, 2014.
[17] C. Setz, et al., “Discriminating Stress from Cognitive load using a Wearable EDA Device”, IEEE Transactions on
Information Technology in Biomedicine, vol. 14, no. 2, pp. 410-417, 2010.
[18] R. Guo, et al., “Pervasive and Unobtrusive Emotion Sensing for Human Mental Health”, Pervasive Health '13
Proceedings of the 7th International Conference on Pervasive Computing Technologies for Healthcare,
pp. 436-439, 2013.
[19] A. Nakasone, et al., “Emotion Recognition from Electromyography and Skin Conductance”, Proc. of the 5th
International Workshop on Biosignal Interpretation, pp. 219-222, 2005.
[20] M. Stikic, et al., “EEG-based Classification of Positive and Negative Affective States”, Brain-Computer Interfaces,
vol. 1, no. 2, pp. 99-112, 2014.
[21] Y. Lu, et al., “Combining Eye Movements and EEG to Enhance Emotion Recognition”, Proceedings of the Twenty-
Fourth International Joint Conference on Artificial Intelligence (IJCAI), 2015.
[22] M. V. Villarejo, et al., “A Stress Sensor based on Galvanic Skin Response (GSR) Controlled by ZigBee”, Sensors,
vol. 12, no. 5, pp. 6075-6101, 2012.
[23] K. Khalfallah, et al., “Non invasive Galvanic Skin Sensor for Early Diagnosis of Sudomotor Dysfunction:
Application to Diabetes”, IEEE Sensor J., vol. 12, pp. 456-463, 2010.
[24] K. Setz, et al., “Towards Long Term Monitoring of Electrodermal Activity in Daily Life”, Proceedings of 5th
International Workshop on Ubiquitous Health and Wellness, Copenhagen, Denmark, 2010.
[25] D. Liu and M. Ulrich, “Listen to Your Heart: Stress Prediction using Consumer Heart Rate Sensors”, CS229
Machine Learning, Autumn 2013.
[26] V. N. Vapnik, “The Nature of Statistical Learning Theory”, Springer-Verlag New York, Inc., 1995.
[27] C. Campbell and Y. Ying, “Learning with Support Vector Machines”, Morgan & Claypool Publishers, 2011.
[28] J. R. Crawford and J. D. Henry, “The Positive and Negative Affect Schedule (PANAS): Construct validity,
measurement properties and normative data in a large non‐clinical sample”, British Journal of Clinical Psychology,
vol. 43, no. 3, pp. 245-265, 2004.
[29] S. B. Gunawan, “Deteksi Status Emosi Manusia Dengan Algoritma Support Vector Machine Menggunakan Data
Galvanic Skin Response”, Theses, Universitas Atma Jaya Yogyakarta, 2017.
[30] D. Grimes, et al., “Feasibility and Pragmatics of Classifying”, CHI 2008 Proceedings: Cognition, Perception, and
Memory, 2008.
[31] X. Li and B. Marlin, “Classification of Sparse and Irregularly Sampled Time Series with”, Proceeding UAI'15
Proceedings of the Thirty-First Conference on Uncertainty in Artificial Intelligence, pp. 484-493, 2015.
[32] I. B. Widodo and B. Widjaja, “Automatic Lag Selection in time series Forecasting using Multiple Kernel
Learning”, International Journal of Machine Learning and Cybernetics, vol. 7, no. 1, pp. 95-110, 2016.
[33] E. T. Solovey, et al., “Classifying Driver Workload using Physiological and Driving Performance Data: Two Field
Studies”, Proceeding CHI '14 Proceedings of the SIGCHI Conference on Human Factors in Computing Systems,
pp. 4057-4066, 2014.
[34] A. Ambarwari, et al., “Biometric Analysis of Leaf Venation Density Based on Digital Image”, Telecommunication
Computing Electronics and Control, vol. 16, no. 4, 2018.
[35] R. A. v. d. Berg, et al., “Centering, Scaling, and Transformations: Improving the Biological Information Content of
Metabolomics Data”, BMC Genomics, 2006.
[36] K. Daimi and S. Banitaan, “Using Data Mining to Predict Possible future Depression cases”, International Journal
of Public Health Science (IJPHS), vol. 3, no. 4, pp. 231-240, 2014.
BIOGRAPHIES OF AUTHORS
Djoko Budiyanto Setyohadi received the B.E. degree in Electrical Engineering from Universitas
Gadjah Mada, Yogyakarta, Indonesia in 1990, the M.Eng. degree in Computer Science, Information
Management from the Asian Institute of Technology, Bangkok, Thailand, in 1998 and the Ph.D.
degree in Computer Science from University Kebangsaan Malaysia. He is an Associate Professor of
Informatics Engineering at the Faculty of Industrial Engineering, Universitas Atma Jaya Yogyakarta.
His current research interests include Information System, Human Computer Interface and Data
Engineering.

 ISSN: 2088-8708
4014
Sri Kusrohmaniah received the B.E. degree in Psychology from UniversitasGadjahMada,
Yogyakarta, Indonesia while her Ph.D. degree in Psychology from University Kebangsaan Malaysia,
in 2017. She is a senior lecturer in Psychology Department, Gadjah Mada University Yogyakarta.
Her current research interests include Neuroscience, Human Computer Interface and Experimental
Psychology.
Sebastian Bagya Gunawan received the Bachelor Degree in 2017 from University of Atma Jaya
Yogyakarta majoring in Information System. He is a research assistant in Data Engineering and
InformationSystem Research Group. His current research interests include Human Computer
Interface and Data Engineering.
Pranowo received the B.Eng. Degree in Mechanical Engineering from UniversitasGadjahMada,
Indonesia, in 1996, the M.Eng. degree in electrical engineering from the Universitas Gadjah Mada,
Indonesia, in 2002 and Ph.D. degree in Electrical Engineering and Information Technology from
Universitas Gadjah Mada, Indonesia, in 2010. He is currently a Senior Lecturer in Department of
InformaticEngineering with the Faculty of Industrial Technology, Universitas Atma Jaya
Yogyakarta, Indonesia. His research interests include numerical simulation and modeling, computer
vision and parallel programming.
Anton Satria Prabuwono is currently a Professor in Faculty of Computing and Information
Technology in Rabigh, King Abdulaziz University, Saudi Arabia. He joined Faculty of Information
Science and Technology, University Kebangsaan Malaysia (UKM) in 2009. He then joined Faculty
of Computing and Information Technology in Rabigh, King Abdulaziz University in 2013. He
served as Editor, Technical Committee, and Reviewer in many International Journals and
Conferences. He is a senior member of IEEE and a member of ACM. In general, his research
interests include computer vision, intelligent robotics, and autonomous systems.

Galvanic Skin Response Data Classification for Emotion Detection

More Related Content

What's hot (20)

Similar to Galvanic Skin Response Data Classification for Emotion Detection (20)

More from IJECEIAES (20)

Recently uploaded (20)

Galvanic Skin Response Data Classification for Emotion Detection