Access to this full-text is provided by Wiley.
Content available from Computational Intelligence and Neuroscience
This content is subject to copyright. Terms and conditions apply.
Research Article
Recognition of the Impulse of Love at First Sight Based on
Electrocardiograph Signal
Jin Zhang ,
1
Guangjie Yuan,
2
Huan Lu,
3
and Guangyuan Liu
1
,
2
,
3
1
College of Electronic and Information Engineering, Southwest University, Chongqing, China
2
Institute of Affective Computing and Information Processing, Southwest University, Chongqing, China
3
Chongqing Key Laboratory of Nonlinear Circuits and Intelligent Information Processing, Southwest University,
Chongqing, China
Correspondence should be addressed to Guangyuan Liu; liugy@swu.edu.cn
Received 10 December 2020; Revised 19 February 2021; Accepted 10 March 2021; Published 24 March 2021
Academic Editor: Fivos Panetsos
Copyright ©2021 Jin Zhang et al. is is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
e impulse of love at first sight (ILFS) is a well known but rarely studied phenomenon. Despite the privacy of these emotions,
knowing how attractive one finds a partner may be beneficial for building a future relationship in an open society, where partners
are accepting each other. erefore, this study adopted the electrocardiograph (ECG) signal collection method, which has been
widely used in wearable devices, to collect signals and conduct corresponding recognition analysis. First, we used photos to induce
ILFS and obtained ECG signals from 46 healthy students (24 women and 22 men) in a laboratory. Second, we extracted the time-
and frequency-domain features of the ECG signals and performed a nonlinear analysis. We subsequently used a feature selection
algorithm and a set of classifiers to classify the features. Combined with the sequence floating forward selection and random forest
algorithms, the identification accuracy of the ILFS was 69.07%. e sensitivity, specificity, F1, and area under the curve of the other
parameters were all greater than 0.6. e classification of ECG signals according to their characteristics demonstrated that the
signals could be recognized. rough the information provided by the ECG signals, it can be determined whether the participant
possesses the desire to fall in love, helping to determine the right partner in the fastest time; this is conducive to establishing a
romantic relationship.
1. Introduction
e impulse of love at first sight (ILFS) is a significant initial
attraction [1], that is, a strong desire to relate with another
person, and is a complex phenomenon that includes eval-
uation, appreciation, and subjective experience of physio-
logical changes. ILFS can be observed in many literary and
artistic works. In real life, the concept of ILFS is accepted by
most people. For example, approximately one-third of
westerners report that they have experienced ILFS [2].
Moreover, studies have observed that ILFS can affect rela-
tionships [3, 4]. A relationship between couples involving
the ILFS is more passionate, causing the relationship to be
more stable and satisfying [5]. Vico et al. [6] observed that, in
the presence of a favorite face, heart rate and skin con-
ductance activity increase, along with valence and arousal
and a reduction in dominance evaluation. Fisher [7] found
that the psychological responses of ILFS include excitement,
increased energy, tremor, rapid heartbeats, and shortness of
breath. Nevertheless, almost no research has been conducted
on recognizing the ILFS. In general, the ILFS is a type of
emotional state that can be studied by referring to previous
methods of emotion recognition.
In recent years, physiological signals, such as electro-
encephalograms [8], electrocardiograms (ECGs) [9–11],
electromyography [12], photoplethysmography [13], gal-
vanic skin [14] response, and respiration, have been widely
applied in the field of emotion recognition. On the one hand,
behavioral data (such as facial expressions and body pos-
tures) and voice data are easily manipulated by subjective
consciousness [15]; on the other hand, physiological signals
are real-time and continuous signals that can be used to
Hindawi
Computational Intelligence and Neuroscience
Volume 2021, Article ID 6631616, 9 pages
https://doi.org/10.1155/2021/6631616
better analyze the expression and conversion between dif-
ferent emotional states. Among these physiological signals,
emotion recognition using ECG signals has become an
important topic in the field of emotion computing. First,
ECG signal-derived features, such as heart rate (HR) and
heart rate variability (HRV), have been observed as reliable
physiological indicators of emotion recognition [16, 17]. For
example, Kreibig [18] demonstrated that happiness results in
a reduction in HRV while joy and entertainment increase
HRV. Research by Lichtenstein et al. [19] indicated that
there are significant differences in HRV corresponding to
anger and happiness, anger and satisfaction, and sadness and
happiness. In addition, Rainville [20] demonstrated that HR
and HRV characteristics can be used to distinguish four
emotions: anger, fear, happiness, and sadness. Second, ECG
signals have been widely used for emotion recognition owing
to the low cost, portability, wearability, and wireless ad-
vantages of ECG devices. Karthikeyan et al. [21] distin-
guished between relaxed and stressed states using ECG
signals and achieved a classification accuracy of 94.6%. Guo
et al. [22] extracted HRV features from ECG signals and used
support vector machines (SVMs) to classify different
emotional states. e results demonstrated that the two
emotional states (positive/negative) attained 71.4% accuracy.
Castaldo et al. [23] evaluated the potential of stress detection
using an ultra-short-term HRV analysis. e experimental
results showed that the sensitivity, specificity, and accuracy
of classification surpassed 60% using ultra-short-term HRV
features for classification. Hsu et al. [15] proposed an ECG-
based automatic emotion recognition algorithm. e clas-
sification accuracy of positive/negative valence, high/low
arousal, and three types of emotions (joy, sadness, and
peacefulness) using a least-squares SVM were 82.78%,
72.91%, and 61.25%, respectively.
In addition, in the study of emotion recognition, pictures
[24], music [25], movies [26], and text [27] are frequently
used to elicit emotions. is study examines ILFS when two
people meet each other. Conducting a speed dating scenario
with hundreds of participants in a laboratory environment is
not feasible. Moreover, the ILFS studied in this study can be
generated in a very short time. erefore, in this study, we
used images to induce ILFS and used ECG signals to classify
and recognize ILFS. Also, we designed an accurate experi-
ment to collect ECG signals from participants during the
viewing period. Subsequently, we developed an automatic
ILFS recognition algorithm to detect the Rwave, generate
important features related to the ILFS, and effectively
identify the ILFS.
e remainder of this paper is organized as follows:
Section 2 describes the experimental setup and protocol. e
proposed ECG-based ILFS recognition algorithm is intro-
duced in Section 3. Section 4 presents the results and cor-
responding discussion. Section 5 presents the conclusions of
this study.
2. Experimental Setup
2.1. Experiment Material. In this study, various factors were
comprehensively considered to select photos as the stimulus
material; 800 photos of smiling men and women were
purchased and downloaded from a photo website. Subse-
quently, these photos were cropped into bust photos with
uniform properties, for example, size, brightness, and
resolution.
Unified processed pictures were scored and formal test
materials were selected. Psychologists have shown that
facial attractiveness is strongly linked to ILFS [28, 29].
Every time the unit of attraction increases by one level, the
likelihood of the ILFS will increase by a factor of nine.
erefore, 60 college students (30 men and 30 women) with
no colorblindness or physical/mental health were recruited
from Southwestern University to evaluate the facial at-
tractiveness of photos of the opposite sex and were asked to
subjectively evaluate facial attractiveness on a scale of 1
(not at all) to 9 (extremely). We then selected 240 male and
240 female photos from those evaluated as material that
induced ILFS (high attraction: average: low
attraction ≈0.25 : 0.6 : 0.15).
2.2. Participants. e researchers recruited 46 healthy
Southwestern University students (24 women and 22 men;
mean age, 19.7 ±1.6 years). e participants were required to
abstain from vigorous exercise for 2 h before the experiment
to avoid a rapid heart rate, which would affect the experi-
mental data and results. However, owing to equipment
problems, the data of the three students were not used.
All participants provided written informed consent.
Before data collection, all methods were approved by the
Human Ethics Research Committee of Southwestern
University.
2.3. Experimental Context. is experiment was divided into
two sessions (two sessions were performed at least one day
apart). Each session contained 120 stimuli. Each session had
two blocks and each block contained 60 stimulus materials.
In the experiment, the presentation time of each stimulus
material was 10 s and the participants were evaluated
according to their emotional state. After each block, a
neutral landscape and a piece of light music were presented
for 4 min. e experimental paradigm is illustrated in
Figure 1.
At the beginning of the experiment, the subjects sat
quietly in a chair and their bodies were in a state of natural
relaxation. e corresponding picture stimulus materials
were then presented according to the written emotion-in-
duced experimental paradigm to induce the ILFS. After the
subjects watched the stimulus materials, they performed the
emotion induction evaluation and subjectively reported the
ILFS induction intensity for each stimulus material, in the
range of 0 (none) to 3 (extreme). ECG signals were collected
using an MP150 system and the sampling frequency was set
to 1000 Hz. After the experiment was completed, the subjects
were asked to look at the pictures again and subjectively
report their arousal, valence, dominance, and attraction, in
the range of 1 to 7. e self-report rating scale used here was
a Likert table [30].
2Computational Intelligence and Neuroscience
3. Methodology
In summary, ECG signals were recorded for 46 participants
observing 240 pictures of the opposite sex. Subsequently, the
ECG signals were preprocessed to remove the interference
and noise. After noise removal, feature extraction was
performed on the signals and the time domain, frequency,
and nonlinear features of the ECG signal were extracted.
After extracting some statistical features (indices), we
employed a feature selection algorithm to reduce the feature
dimensions, thereby reducing the computational cost. Fi-
nally, different classifiers are used for sentiment classifica-
tion. e frame diagram of the state recognition of the ILFS
is shown in Figure 2.
3.1. Preprocessing. Before preprocessing, the ECG signal was
downsampled to 200 Hz.
e ECG signal is a nonstationary weak signal that easily
receives interference from itself and the outside environ-
ment; this interference and noise may conceal useful in-
formation. Before feature extraction, the original ECG signal
must be preprocessed. ECG frequently includes baseline
drift below 1 Hz, power frequency interference at 50Hz, and
electromyographic interference. During preprocessing, a
discrete wavelet transform—a common method for re-
moving noise [31]—was used. e original ECG signal was
scaled using a discrete wavelet transform, the approximate
coefficients and detail coefficients of each layer were
extracted, and the soft threshold function was used to
process the detail coefficients. Subsequently, a pure ECG
signal was reconstructed.
e noise-removed ECG signal was divided into 10 s
time signals when the stimulus material appeared as the
starting point. Subsequently, the Pan–Tompkins peak de-
tection algorithm was used to locate the R-wave peak to
obtain the RR interval [32]. HRV parameters can be ob-
tained through feature extraction of the RR interval. HRV is
a reliable marker of activity in the autonomic nervous system
and reflects the time change of a continuous heartbeat [33].
3.2. Feature Extraction. In this study, twenty-five features
were extracted from the ECG signals, including the HRV
time domain, frequency domain, and nonlinear character-
istics; details of the feature information are presented in
Table 1.
3.3. Construction of ILFS and Non-ILFS Datasets. Before
constructing the datasets, we first removed the abnormal
data, which would have affected the classification results.
e median absolute deviation (MAD) algorithm can
effectively remove outliers from the data [34]. e MAD and
outlier removal methods are shown in the following
equations, respectively:
MAD �medianixi−medianjxj
,(1)
xi≤median xi
−5×MAD,
xi≥median xi
+5×MAD,
(2)
where xjis one of the nsample values and medianiis the
median of the series.
60 stimuli
60 stimuli
60 stimuli
60 stimuli
Rest
Rest
Session 1 (120 stimuli)
Session 2 (120 stimuli)
Average High attraction Low attraction
Figure 1: Experimental paradigm.
Computational Intelligence and Neuroscience 3
By summarizing previous studies on the ILFS
[6, 7, 35], we consider that the ILFS exhibits the char-
acteristics of high arousal, high price, high attractiveness,
and high dominance. erefore, combining the two
evaluations in the experiment, the data with high arousal,
high price, high attractiveness, and high dominance were
screened from the ILFS data (data with a level of 1 for the
ILFS were not used) as the dataset of ILFS states. In
addition, for the non-ILFS data, data with low arousal, low
valence, low attractiveness, and low dominance were
selected from the data without the ILFS as the non-ILFS
dataset.
3.4. Feature Selection. A feature selection algorithm can
remove redundant features and reduce the quantity of data,
thereby improving the classification accuracy and signifi-
cantly reducing the computational cost [36]. We thus
Table 1: ECG characteristic description.
Number Symbol Feature description
Time-domain features
1 Mean_RR Mean of RR intervals
2 CVRR e coefficient of variance of RR intervals
3 SDRR Standard deviation of RR intervals
4 RMSSD Root mean square of successive differences of RR intervals
5 MSD Mean of the absolute values of the first differences of RR intervals
6 SDSD Standard deviation of successive differences of RR intervals
7 NN50 Number of interval differences of successive RR intervals greater than 50 ms
8 PNN50 Corresponding percentage of RR50
9 NN20 Number of interval differences of successive RR intervals greater than 20 ms
10 PNN20 Corresponding percentage of RR20
11 Mean_HR Average heart rate
12 QD Quartile deviation of RR intervals
Nonlinear features
13 SD1 Standard deviation for Tdirection in Poincare plot
14 SD2 Standard deviation for Ldirection in Poincare plot
15 SD1_SD2 SD1/SD2
16 CSI Cardiac sympathetic index
17 CVI Cardiac vagal index
18 modified_CSI Modified CSI
19 LZC LZ complexity
Frequency-features
20 TP Power of range 0.04–0.4 Hz of the PSD of RR intervals
21 LF Power of range 0.04–0.15 Hz of the PSD of RR intervals
22 HF Power of range 0.15–0.4 Hz of the PSD of RR intervals
23 LF/HF Proportion of LF to HF
24 nLFP Proportion of LF to LF + HF
25 nHFP Proportion of HF to LF + HF
Picture
ECG sensor
Collection ECG signal
Data preprocessing
Waveform detection
Feature extraction
Feature selection
Classication
model
e impulsion of love
recognition
(a) Data acquisition
(b) Data preprocessing
and waveform detection
(c) Feature processing (d) Emotion recognition
Figure 2: Frame diagram of state recognition of ILFS.
4Computational Intelligence and Neuroscience
selected the sequence floating forward selection (SFFS) al-
gorithm. e SFFS algorithm selects an optimal feature
subset as the classification input and can solve the local
optimization problem of the feature set to a certain extent
[37].
SFFS combines sequential forward selection (SFS) and
sequential backward selection (SBS) algorithms. e SFFS
has three parts: insertion, conditional exclusion, and
termination.
First, let Fk�fi:1≤i≤k
be a feature subset com-
posed of kfeatures selected from the original feature set
Y�yi:1≤i≤n
, where nis the total number of features.
e evaluation function of the optimal feature subset was
J(·).
Step 1. Inclusion: beginning from the empty set F�∅,
use the SFS method to select the most important feature
f+from Y−Fk
and Fkto form a new feature subset
Fk+1, and Fk+1�Fk+1+f+. Set k�k+ 1 to execute Step
2.
Step 2. Conditional exclusions determine the most
important feature (f−) from Fk+1, if f−is the most
important feature in Fk+1, and J(Fk+1−f−)>J(Fk);
delete f−from Fk+1to form a new feature subset Fk
′and
Fk
′�Fk+1−f−. ereafter, Step 3 is performed. In
addition, if J(Fk+1−f−)<J(Fk), return to Step 1.
Step 3. Termination. Set k�k– 1; if kis equal to the
expected number of features, stop. Otherwise, set
Fk�Fk
′,J(Fk) � J(Fk
′), and return to Step 1.
In this study, two nested 10-fold cross-validation
schemes were used to obtain reliable model estimates for
feature selection and model training [35]. e best feature
subset is selected in the inner loop. In the outer loop, using
the selected best feature subset, the classifier was evaluated
using 10-fold cross-validation.
4. Results and Discussion
is section presents a series of results (feature analysis and
classification results) to evaluate the effectiveness of the
proposed approach. In addition, the results were compre-
hensively discussed.
4.1. Feature Analysis. We evaluated whether the charac-
teristics of the ECG are shown in Table 1; the characteristics
of the ILFS data sample, and the characteristics of the non-
ILFS data were significantly different. e Wilcoxon signed-
rank test is the most extensive nonparametric rank-sum test
method for two independent groups [38]. A pvalue of less
than 0.05 indicates that a significant difference exists be-
tween the ILFS and non-ILFS. As Figure 3 illustrates, the
results of the Wilcoxon test indicated that the differences in
ILFS and non-ILFS status for features #3, #7, #8, #9, #12, #17,
#20, and #21 were insignificant. Consistent with the results
in [6, 7], the Wilcoxon test results of feature #11 indicate that
the ILFS state exhibits a higher heart rate.
Although some features are not significantly different
between the ILFS and non-ILFS states, the classification
performance can be significantly improved when used in
combination with other features [39]. erefore, we used 25
heartbeat feature vectors to represent each sample in the
ILFS and non-ILFS state datasets.
Feature selection involves selecting the fewest features
without affecting the classification effect. us, in this study,
10-fold cross-validation schemes based on the SFFS algo-
rithm were used for feature selection. e number of features
was changed from 1 to 25 for training and the best feature
subset was selected. Figure 4 shows the accuracy of the five
classifiers for selecting different numbers of features. e
features corresponding to the maximum accuracy of the
different classifiers were used as the optimal feature subset of
the classifier.
Table 2 lists the best feature subsets of the different
classifiers. It can be seen from Table 2 that feature #1 is one
of the best performing features in each classifier and feature
#1 is reduced in the ILFS state. Consistent with the liter-
ature results [7, 35], ILFS produces physiological reactions,
for example, excitement, a rapid heartbeat, an increased
heart rate in the excited state, and a reduced average RR
interval.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
p value
Features
0.05
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Figure 3: Characteristic pvalue between the ILFS and non-ILFS.
Computational Intelligence and Neuroscience 5
4.2. Classification Result. In this research, the ILFS and non-
ILFS samples were classified using a set of widely used
classifiers, such as SVM, random forest (RF), and naive
Bayes (NB). In addition, sensitivity (Se), specificity (Sp), F1-
score (F1), area under the curve (AUC), accuracy (ACC),
and other parameters were used to evaluate the performance
of the classification scheme. Table 3 and Figure 5 present the
classification performance of five classifiers without feature
selection for ECG signals. Among these classifiers, RF ex-
hibits the best classification accuracy, with a result of 66.04%.
Other classifiers recognized the ILFS and their classification
accuracy was approximately 60%. e parameters Se, Sp, F1,
and AUC of the classifier were all approximately 0.6.
During the analysis presented in the previous section, the
optimal feature subset of the classifier was obtained based on
the SFFS algorithm and the optimal feature subset was used
to evaluate the classifier using 10-fold cross-validation.
Table 4 and Figure 6 show the classification performance of
the five classifiers after feature selection. e results dem-
onstrate that the highest accuracy rate of 69.07% is obtained
for the classifier RF and features #1, #3, #8, #12, and #24
constitute the best feature subset; the parameters Se, Sp, F1,
and AUC of the classifier RF were better than those of the
other classifiers; however, the parameters of all classifiers are
greater than 0.6, indicating that ILFS can be classified and
recognized.
Figure 7 shows that after using feature selection, the
classification effect of 5 classifiers is improved. In addition, it
can be seen that the feature selection method (combined
with the RF classifier) is optimal for identifying ILFS
emotions, compared to other machine learning algorithms.
In previous studies, few researchers have examined the
mapping pattern between the ILFS and physiological signals.
erefore, in this paper, a study on the classification and
recognition of the ILFS based on ECG signals is proposed,
that is, the use of an ECG signal to identify whether someone
0.5
0.52
0.54
0.56
0.58
0.6
0.62
0.64
0.66
0.68
0.7
Accuracy
Number of features
12345678910111213141516171819202122232425
RF
KNN
SVM
NB
DT
Figure 4: Results of the accuracy of five classifiers for selecting different numbers of features.
Table 2: e best feature subsets of five classifiers.
Classifier Selected features
SVM 1, 3, 8, 15, 14
RF 1, 3, 8, 12, 24
NB 1, 2, 11
KNN 4, 7, 8, 12
DT 1, 8, 24
Table 3: Classification performance of the five classifiers without
feature selection.
Classifier Se Sp F1 AUC ACC
SVM 0.7103 0.5305 0.6512 0.6209 0.6221
RF 0.6984 0.6186 0.6717 0.6616 0.6604
NB 0.5 0.6363 0.5513 0.6017 0.5988
KNN 0.5949 0.6363 0.6037 0.62 0.6128
DT 0.6001 0.6037 0.5977 0.6122 0.6011
SVM: support vector machine; RF: random forest; NB: naive bayes; KNN:
K-nearest neighbor; DT: decision tree.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
SVM
RF
NB
KNN
DT
Se Sp F1 AUC ACC
Figure 5: Classification performance of the five classifiers without
feature selection.
6Computational Intelligence and Neuroscience
is in a state of ILFS. e best classification accuracy rate was
66.04% for all signal characteristics. With the SFFS feature
selection algorithm, the best classification accuracy in-
creased to 69.03%. e experimental results show that ILFS
can be classified and identified based on ECG signals;
however, the recognition and classification of the ILFS based
on ECG signals are not very accurate. e following may be
factors that affect the classification of ILFS:
(1) e recognition effect of the ILFS is related to the
classifier used. Selecting a more advanced classifica-
tion algorithm can improve the classification effect.
(2) e ILFS is highly related to the subjects’ aesthetic
preferences and the emotional intensity induced by
the selected stimulus photos is insufficient.
(3) e ILFS is a complex emotional state. Accurately
reflecting the changes in the ILFS using only an ECG
signal is difficult.
erefore, in future research, for better classification and
recognition of the ILFS, it is necessary to (1) determine a
more advanced classification algorithm, (2) use different
stimuli (e.g., video) to induce a higher intensity of the ILFS,
and (3) use a variety of physiological signals.
5. Conclusions
is study attempted to identify the ILFS based on ECG
signals. Our research demonstrated that the ILFS is sepa-
rable. Based on the recognition of the ILFS using the ECG,
through the information provided by the physiological
signal, people can determine whether they have the ILFS.
Determining the right partner in the fastest time is con-
ducive to establishing a relationship. Moreover, owing to the
advantages of low cost, portability, and wearable devices, the
ECG signal-based ILFS recognition algorithm can be
combined with wearable devices, which can better match the
cardiac target in specific scenarios, online or offline.
Data Availability
e data used to support the findings of this study are
available from the corresponding author upon request.
Conflicts of Interest
e authors declare that they have no conflicts of interest.
Acknowledgments
e authors thank all the subjects and the experimenters
participating in the experiment. is work was supported in
part by the National Natural Science Foundation of China
(nos. 61472330 and 61872301).
References
[1] F. Zsok, M. Haucke, C. Y. De Wit, and D. P. H. Barelds, “What
kind of love is love at first sight? An empirical investigation,”
Personal Relationships, vol. 24, no. 4, pp. 869–885, 2017.
[2] E. Naumann, “Love at first sight: the stories and science
behind instant attraction,” Frontiers in Psychology, vol. 7,
2001.
[3] L. Custer, D. Holmberg, K. L. Blair, and T. Orbuch, So How
Did You Two Meet?” Narratives of Relationship Initiation:
Handbook of Relationship Beginnings, Lawrence Erlbaum
Associates, Mahwah, NJ, USA, 2008.
[4] B. Fehr, Love: Conceptualization and Experience, American
Psychological Association, Worcester, MA, USA, 2015.
[5] N. Alea and S. C. Vick, “e first sight of love: relationship-
defining memories and marital satisfaction across adulthood,”
Memory, vol. 18, no. 7, pp. 730–742, 2010.
[6] C. Vico, P. Guerra, H. Robles, J. Vila, and L. Anllo-Vento,
“Affective processing of loved faces: contributions from pe-
ripheral and central electrophysiology,” Neuropsychologia,
vol. 48, no. 10, pp. 2894–2902, 2010.
Table 4: Classification performance of five classifiers after feature
selection.
Classifier Se Sp F1 AUC ACC
SVM 0.7684 0.5631 0.6908 0.656 0.6616
RF 0.7194 0.6635 0.6984 0.6903 0.6907
NB 0.627 0.6101 0.6021 0.6194 0.6232
KNN 0.6307 0.631 0.627 0.634 0.6325
DT 0.6343 0.5897 0.6216 0.6209 0.6163
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
SVM
RF
NB
KNN
DT
Se Sp F1 AUC ACC
Figure 6: Classification performance after feature selection.
0.6616
0.6907
0.6232
0.6325
0.6163
0.6221
0.6604
0.5988
0.6128
0.6011
0.54
0.56
0.58
0.6
0.62
0.64
0.66
0.68
0.7
DT
ACC
Feature selection
No feature selection
SVM RF NB KNN
Figure 7: Comparison of the classification accuracy of different
classifiers with no feature selection and feature selection.
Computational Intelligence and Neuroscience 7
[7] H. E. Fisher, A. Aron, D. Mashek, H. Li, and L. L. Brown,
“Defining the brain systems of lust, romantic attraction, and
attachment,” Archives of Sexual Behavior, vol. 31, no. 5,
pp. 413–419, 2002.
[8] R. Nawaz, K. H. Cheah, H. Nisar, and V. V. Yap, “Comparison
of different feature extraction methods for EEG-based
emotion recognition,” Biocybernetics and Biomedical Engi-
neering, vol. 40, no. 3, pp. 910–926, 2020.
[9] G. Valenza, A. Lanata, and E. P. Scilingo, “e role of
nonlinear dynamics in affective valence and arousal recog-
nition,” IEEE Transactions on Affective Computing, vol. 3,
no. 2, pp. 237–249, 2012.
[10] A. O. Akmandor and N. K. Jha, “Keep the stress away with
SoDA: stress detection and alleviation system,” IEEE Trans-
actions on Multi-Scale Computing Systems, vol. 3, no. 4,
pp. 269–282, 2017.
[11] A. Tjolleng, K. Jung, W. Hong et al., “Classification of a
Driver’s cognitive workload levels using artificial neural
network on ECG signals,” Applied Ergonomics, vol. 59,
pp. 326–332, 2017.
[12] B. Cheng and G. Liu, “Emotion recognition from surface
EMG signal using wavelet transform and neural network,”
Journal of Computer Applications, vol. 28, no. 2, pp. 1363–
1366, 2008.
[13] Y. K. Lee, O. W. Kwon, H. S. Shin, J. Jo, and Y. Lee, “Noise
reduction of PPG signals using a particle filter for robust
emotion recognition,” in Proceedings of the 2011 IEEE In-
ternational Conference on Consumer Electronics-Berlin (ICCE-
Berlin), pp. 202–205, Berlin, Germany, September 2011.
[14] W. Wen, G. Liu, N. Cheng, J. Wei, P. Shangguan, and
W. Huang, “Emotion recognition based on multi-variant
correlation of physiological signals,” IEEE Transactions on
Affective Computing, vol. 5, no. 2, pp. 126–140, 2014.
[15] Y.-L. Hsu, J.-S. Wang, W.-C. Chiang, and C.-H. Hung,
“Automatic ECG-based emotion recognition in music lis-
tening,” IEEE Transactions on Affective Computing, vol. 11,
no. 1, pp. 85–99, 2020.
[16] O. Alzoubi, S. K. D’Mello, and R. A. Calvo, “Detecting nat-
uralistic expressions of nonbasic affect using physiological
signals,” IEEE Transactions on Affective Computing, vol. 3,
no. 3, pp. 298–310, 2012.
[17] S. H. Fairclough and C. Dobbins, “Personal informatics and
negative emotions during commuter driving: effects of data
visualization on cardiovascular reactivity & mood,” Inter-
national Journal of Human-Computer Studies, vol. 144, Article
ID 102499, 2020.
[18] S. D. Kreibig, “Autonomic nervous system activity in emotion: a
review,” Biological Psychology, vol. 84, no. 3, pp. 394–421, 2010.
[19] A. Lichtenstein, A. Oehme, S. Kupschick, and T. J¨urgensohn,
Comparing Two Emotion Models for Deriving Affective States
from Physiological Data,” Affect and Emotion in Human-
Computer Interaction, Springer Berlin Heidelberg, New York,
NY, USA, 2008.
[20] P. Rainville, A. Bechara, N. Naqvi, and A. R. Damasio, “Basic
emotions are associated with distinct patterns of cardiore-
spiratory activity,” International Journal of Psychophysiology,
vol. 61, no. 1, pp. 5–18, 2006.
[21] P. Karthikeyan, M. Murugappan, and S. Yaacob, “Analysis of
stroop color word test-based human stress detection using
electrocardiography and heart rate variability signals,” Ara-
bian Journal for Science and Engineering, vol. 39, no. 3,
pp. 1835–1847, 2014.
[22] H. Guo, Y. Huang, C. Lin, J. Chien, K. Haraikawa, and
J. Shieh, “Heart rate variability signal features for emotion
recognition by using principal component analysis and
support vectors machine,” in Proceedings of the IEEE Inter-
national Conference on Bioinformatics & Bioengineering,
pp. 274–277, IEEE, Taichung, Taiwan, October 2016.
[23] R. Castaldo, L. Montesinos, P. Melillo, S. Massaro, and
L. Pecchia, “To what extent can we shorten HRV analysis in
wearable sensing? a case study on mental stress detection,” in
Proceedings of the EMBEC & NBC 2017: Joint Conference of the
European Medical and Biological Engineering Conference
(EMBEC) and the Nordic-Baltic Conference on Biomedical
Engineering and Medical Physics (NBC), Tampere, Finland,
June 2017.
[24] M. Codispoti, M. M. Bradley, and P. J. Lang, “Affective re-
actions to briefly presented pictures,” Psychophysiology,
vol. 38, no. 3, 2001.
[25] J. Kim and E. Andre, “Emotion recognition based on phys-
iological changes in music listening,” IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 30, no. 12,
pp. 2067–2083, 2008.
[26] X.-W. Wang, D. Nie, and B.-L. Lu, “Emotional state classi-
fication from EEG data using machine learning approach,”
Neurocomputing, vol. 129, no. 10, pp. 94–106, 2014.
[27] C. O. Alm, D. Roth, and R. Sproat, “Emotions from text:
machine learning for text-based emotion prediction,” in
Proceedings of the Conference on Human Language, Tech-
nology and Empirical Methods in Natural Language Processing
- HLT ’05, Vancouver, CA, USA, October 2005.
[28] J. L. Sangrador and C. Yela, “‘What is beautiful is loved’:
physical attractiveness in love relationships in a representative
sample,” Social Behavior and Personality: An International
Journal, vol. 28, no. 3, pp. 207–218, 2000.
[29] P. W. Eastwick, L. B. Luchies, E. J. Finkel, and L. L. Hunt, “e
predictive validity of ideal partner preferences: a review and
meta-analysis,” Psychological Bulletin, vol. 140, no. 3,
pp. 623–665, 2014.
[30] R. Likert, “A technique for the measurement of attitudes,”
Archieves of Psychology, vol. 140, no. 22, pp. 1–55, 1932.
[31] H.-Y. Lin, S.-Y. Liang, Y.-L. Ho, Y.-H. Lin, and H.-P. Ma,
“Discrete-wavelet-transform-based noise removal and feature
extraction for ECG signals,” IRBM, vol. 35, no. 6, pp. 351–361,
2014.
[32] J. Pan and W. J. Tompkins, “A real-time QRS detection al-
gorithm,” IEEE Transactions on Biomedical Engineering,
vol. BME-32, no. 3, pp. 230–236, 1985.
[33] P. Ekman, W. V. Friesen, M. O’Sullivan et al., “Universals and
cultural differences in the judgments of facial expressions of
emotion,” Journal of Personality and Social Psychology, vol. 53,
no. 4, pp. 712–717, 1987.
[34] C. Leys, C. Ley, O. Klein, P. Bernard, and L. Licata, “Detecting
outliers: do not use standard deviation around the mean, use
absolute deviation around the median,” Journal of Experi-
mental Social Psychology, vol. 49, no. 4, pp. 764–766, 2013.
[35] S. Pourmohammadi and A. Maleki, “Stress detection using
ECG and EMG signals: a comprehensive study,” Computer
Methods and Programs in Biomedicine, vol. 193, Article ID
105482, 2020.
[36] J. Schenk, M. Kaiser, and G. Rigoll, “Selecting features in on-
line handwritten whiteboard note recognition: SFS or SFFS,”
in 10th International Conference on Document Analysis and
Recognition, ICDAR 2009, Barcelona, Spain, July 2009.
[37] P. Pudil and J. Novoviˇ
cov´
a, Novel Methods for Feature Subset
Selection with Respect to Problem Knowledge,” Feature Ex-
traction, Construction and Selection, pp. 101–116, Springer US,
New York, NY, USA, 1998.
8Computational Intelligence and Neuroscience
[38] F.. , K. S. K. Wilcoxon and R. A. Wilcox, Critical Values and
Probability Levels for the Wilcoxon Rank Sum Test and the
Wilcoxon Signed Rank Test,” Selected Tables in Mathematical
Statistics, pp. 171–259, American Cyanamid, Bridgewater
Township, NJ, USA, 1970.
[39] I. Guyon and A. Elisseeff, “An introduction to variable and
feature selection,” Journal of Machine Learning Research,
vol. 3, no. 6, pp. 1157–1182, 2003.
Computational Intelligence and Neuroscience 9
Available via license: CC BY 4.0
Content may be subject to copyright.
