International Journal of Reconfigurable and Embedded Systems (IJRES)

Vol. 13, No. 1, March 2024, pp. 201~209

ISSN: 2089-4864, DOI: 10.11591/ijres.v13.i1.pp201-209  201

Affective analysis in machine learning using AMIGOS with

Gaussian expectation-maximization model

Balamurugan Kaliappan1, Bakkialakshmi Vaithialingam Sudalaiyadumperumal2,

Sudalaimuthu Thalavaipillai1
1
Department of Computer Science and Engineering, Hindustan Institute of Technology and Science, Padur, India
2
Department of Computing Technologies, SRM Institute of Science and Technology, Kattankulathur, India

Article Info ABSTRACT

Article history: Investigating human subjects is the goal of predicting human emotions in the
real world scenario. A significant number of psychological effects require
Received Jan 4, 2023 (feelings) to be produced, directly releasing human emotions. The
Revised Aug 17, 2023 development of effect theory leads one to believe that one must be aware of
Accepted Sep 5, 2023 one's sentiments and emotions to forecast one's behavior. The proposed line
of inquiry focuses on developing a reliable model incorporating
neurophysiological data into actual feelings. Any change in emotional affect
Keywords: will directly elicit a response in the body's physiological systems. This
approach is named after the notion of Gaussian mixture models (GMM). The
Affective computing statistical reaction following data processing, quantitative findings on emotion
AMIGOS labels, and coincidental responses with training samples all directly impact the
Emotion detection outcomes that are accomplished. In terms of statistical parameters such as
Emotional psychology population mean and standard deviation, the suggested method is evaluated
Machine learning compared to a technique considered to be state-of-the-art. The proposed
system determines an individual's emotional state after a minimum of 6
iterative learning using the Gaussian expectation-maximization (GEM)
statistical model, in which the iterations tend to continue to zero error. Perhaps
each of these improves predictions while simultaneously increasing the
amount of value extracted.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Bakkialakshmi Vaithialingam Sudalaiyadumperumal
Department of Computing Technologies, SRM Institute of Science and Technology
SRM Nagar, Kattankulathur, Chengalpattu District, Tamil Nadu 603203, India
Email: [email protected]

1. INTRODUCTION
The field of sociology has a rich tradition in studying emotions, particularly in the context of devising
viral questions and self-assessment assessments aimed at detecting and understanding human emotional
responses. Human personality predictions with human emotions in real-time scenarios impact the research
industry more. Emotions may be conscious or unconscious; human behavior directly impacts behavior.
Affective computing, influenced by psychological factors, is a difficult field of study with different ideological
paths. Audio signals can accurately depict the emotional impact because pitch variations define the emotional
element. Numerous cross-modal emotion embedding systems employ audio and video correlations inside the
ensemble learning framework to ascertain the genuine emotion elicited by the individual [1].
Long-term mental diseases like depression and anxiety start mildly and gradually affect people's
emotions. Affect sensing is a large and challenging field of study. It is unavoidable to mention the term
"emotion contagion" when explaining the effect scenario. Emotional contagion is a social contagion in which
one person's emotions and behaviors spread to another [2]. The emotional reflection from one person to another

Journal homepage: http://ijres.iaescore.com

202  ISSN: 2089-4864

in a specific circumstance may occur. Actions caused emotions in some cases. When inconvenient, people can
act differently based on how they respond to specific conditions [3].
With the development of artificial intelligence technology, it is now possible to do significant
interactive analysis to understand people's emotions from various perspectives. Standardized datasets are
accessible for research purposes, and the study uses diverse publicly available data. The determination of
emotional affects is achieved by the analysis of speech cues. Pitch and tone alterations serve as evident
indications of emotional transitions or notifications of mood. Neuro-fuzzy logic-based resilience evaluations
are employed to differentiate speech patterns that modify emotional impact [3]. The most obvious expression
of psychological impacts is found on one's face. Expressions are a universally important concept. The mood
can be read from the person's expressions.
The identification of concealed emotions through facial expressions can be challenging under some
circumstances. Another method of expression formulation is to use virtual face vectors and landmark
extractions [4]. By manipulating current neural network challenges, machine learning technology can enhance
algorithms. Using linear discriminant analysis (LDA) models, the extensive collection of feature vectors
acquired from subject analysis data is investigated. Emotion analysis outcomes improve with LDA as trials
increase [5]. The algorithm is proposed for multi-label learning to aid in investigating emotional effects in
several modalities. Diverse modality scenarios allow the emotional effect components to be investigated from
several perspectives regarding the primary polarity of happy and sad emotions [6].
The genuine depiction of the brain's responsiveness to an input is emotion. It expresses the innate
sense present in the predicament. Based on dimensionality, there are two categories of emotion models: 2D is
meant to be two-dimensional models, and 3D is meant to be three-dimensional models. The valance and arousal
dimensions are where the 2D model's most potent emotions can be located. On the other hand, the 3D model
holds valid emotions like valence, arousal dominance, and so forth [7]. The existing research, its drawbacks,
and proposed future research are all summarised in section 2. Studies are underway in the background to
advance section 3. The model selection and design constraint analysis are explained in section 4. The strategy,
data collection, and proposed method are all described in depth in section 5. Challenges faced with the work
and future challenges are addressed in section 6, describing the outcomes and follow-up conversations as
conclusions in section 7.

2. BACKGROUND STUDY
Li et al. [8] multiple polarities concerning emotions are detected using a multi-step enabled deep
emotion detection framework. Deep neural networks (DNN) are used to extract movies and physiological
information from publicly available databases (DNN). Pattern comparison is conducted to analyse and evaluate
the training and testing properties.
Hoang et al. [9] studied detection posture, face, and detection, which is evaluated with a mainstream
multi-task cascaded neural network model using a virtual semantic module. The extraction of the reasoning
stream is accomplished by the utilisation of a multi-level perceptron (MLP). The utilisation of the EXOTIC
dataset, which incorporates simulated heat stream patterns, enhances the efficacy of the detection technique.
Islam et al. [10] stated that emotion detection is a method for identifying and extending a person's
emotional state. Upon the basis of in-depth and surface-level learning, the detection and evaluation of irrational
emotions is put into practice. The coupling of electrocardiogram (ECG) and electroencephalogram (EEG) data
is utilised to showcase the interconnectedness in the context of emotion identification. The PRISMA technique
is employed for comprehensive analysis, encompassing the processes of identification, screening, and
eligibility assessment.
Albraikan et al. [11] present study showcases a system that utilises the MAHNOB dataset and the K-
nearest neighbour method for the purpose of analysing emotion through the application of weighted multi-
dimensional discrete wavelet transform (DWT). Following a series of training rounds, a meta classifier utilises
a combination of video clips depicting various emotions to determine the ultimate emotional affect. The present
study used MWDWT simulations to identify and delineate nine distinct emotional states. The user's text does
not contain any information to rewrite in an academic manner [11].
Qayyum et al. [12] explained a method of emotion recognition via an Android application is offered
due to the prevalence of mobile use. Convolutional neural networks (CNN) and recurrent neural networks
(RNN) are combined to generate a powerful model for emotion detection. CNN and RNN have both accuracy
rates of 65% and 41%, respectively. The mentioned recommendation mechanism is used for fresh content.
Bakkialakshmi and Sudalaimuthu [13] in a self-supervised learning system, unlabeled data are
converted to bias weights based on iterative learning and used return updates. The hardware sensors collected
with diverse temporal features assess an ECG-based emotion identification system. The maximum accuracy
achieved using the typical emotion datasets of AMIGOS, SWELL, WESAD, and DREAMER was 97%.

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 1, March 2024: 201-209
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  203

2.1. Scholarly articles

For algorithm selection, many existing implementations are explored. The middle ground between
supervised and unsupervised learning strategies is self-supervised learning models. The advantage of the self-
supervised learning strategy is that it allows for learning large amounts of unlabeled data. The biased weights
are constantly changed due to the unlabeled data, allowing for the downstream version of raw data [14]. Deep
belief networks are a reliable way of understanding complex data relationships. Complex structures are
required for multi-feature analysis models, which rely on unique data pairings [15].
CNN methods are automatic feature mapping blocks that can be tweaked for more detailed analysis.
Changing the CNN's preceding layers can build an adaptive network. Wholly connected, ReLu and Max-
pooling layers are a few samples of feature selection blocks that can be improved to create adaptive designs
[15]. Gaussian mixture models (GMM) are utilised to characterise probabilistic data originating from the finite
Gaussian distribution within a random space. The model utilises a process of consolidating the relevant data
into a structured grouping in order to improve the accuracy of the regression analysis [16].

2.2. Datasets available

One-minute-gradual (OMG)-emotion-behaviors dataset: the 12,567 YouTube videos that make up the
OMG-emotion-behaviors dataset have an average duration of one minute. The videos are categorised according
to several emotional states, namely joy, sadness, surprise, fright, and disgust. The dataset comprises a collection
of standard one-minute videos that evoke the aforementioned sensation. The OMG emotion dataset [17] was
one of the established models utilised for eliciting emotions. The MAHNOB-HCI dataset incorporates a
module for emotion recognition based on keyword tagging, as opposed to utilising an emotion rating system.
The dataset has been structured according to a group of 24 individuals who were selected as volunteers. These
volunteers were exposed to a total of 20 unique videos that were designed to elicit brain stimulation and
subsequently suppress their genuine emotional responses [18]. The proposed approach aims to mitigate the
limitations associated with the presence of identical polarity in decision-making processes, with the ultimate
goal of enhancing the accuracy of predictions. Additionally, the methodology involves the evaluation of an
ensemble algorithm based on a GMM.

3. METHOD
3.1. Data collection
The AMIGOS data collection serves as a widely used resource for conducting research on emotional
personality and mood. It has data pertaining to both individuals and groups, which have been annotated
externally and characterised by their personality profiles. Neurophysiological recordings from the subject
during the exam include ECG, EEG, and galvanic skin response (GSR) signals [19]. The volunteers are shown
short and long video experimental movies during the test. Forty volunteers saw 16 successful movies that
elicited feelings such as valence, arousal dominance, familiarity, and like in the brain. The viewers experience
a range of fundamental emotions when watching the videos, encompassing neutrality, happiness, sadness,
surprise, fear, anger, and contempt. The determination of mood assessment should be made by considering the
available information and evaluating the patients' kinematics. AMIGOS is a well-documented dataset that has
passed a self-assessment exam. The AMIGOS dataset considers ECG, EEG, and GSR signals in the proposed
system.

3.2. Gaussian expectation-maximization algorithm

The probability clustering model is insufficient in its ability to effectively categorise the given data
into similar groups. However, the expectation-maximization technique relies on the GMM as its foundation
[20]. The predetermined categories' density and each observation validate a set of classes. Data that is
concentrated and belongs to the same class can be grouped by it. The Gaussian expectation-maximization
(GEM) algorithm is an iterative technique used to iteratively estimate maximum-likelihood values for model
parameters in scenarios when the available data is inadequate, includes misclassified data feature points, or
involves unique eigen variables. GEM analyses a new dataset by assigning arbitrary values to the missing data
points. By adding in the missing points, those new values are applied to train the model and the recursively
find better covariate data [21]. The standard distribution analysis, explained below, is first step in the procedure.

3.3. Normal distribution

The expectation-maximization procedure in a Gaussian space allows for the random location selection
of covariate points from the input data. Iterative loops are commonly employed to continuously search for
further data derived from statistical measurements, encompassing the standard deviation, variance, and
population mean of the established pattern. In general, the normal distribution is given by (1).

Affective analysis in machine learning using AMIGOS with Gaussian … (Balamurugan Kaliappan)
204  ISSN: 2089-4864
1 𝑥−𝜇
1
(𝑥) = 𝑒−2( 𝜆
)
(1)
𝜆√2𝜋

Where,
−𝛼<𝑥<𝛼,
𝜆→ 𝑉𝑎𝑟𝑖𝑒𝑛𝑐𝑒,
𝜇→ 𝑀𝑒𝑎𝑛 𝑜𝑓 𝑡ℎ𝑒 𝑝𝑜𝑝𝑢𝑙𝑎𝑛𝑎𝑡𝑖𝑜n

4. SYSTEM ARCHITECTURE
4.1. Design architecture
Figure 1 demonstrates the system architecture and analysis of the proposed GEM model, built on the
normal probability distribution. The input data include preprocessed physiological signal records, including
ECG, EEG, and GSR [22]. The emotional effect that causes the variances makes these patterns distinctive.
According to the stated test record, the impact point is prohibited. Only the correlated points may be obtained
using the normal distribution of random data of overall physiological information.

Figure 1. GEM model

4.2. Summary of implementations

The AMIGOS dataset consists of a succession of large-scale values obtained from ECG, EEG, GSR,
and self-reports on the emotional affects of volunteers [23]. There are two aspects to the suggested Emo-GEM
paradigm. The first stage preprocesses and rearranges the processed data from the AMIGOS into sample
frames. These samples are fed into the GEM method, which analyzes the covariate points in the data and
calculates lambda (λ) and sigma (Σ) values. The model is tested repeatedly to achieve lower latency and a 0%
error rate. The adaptive weights will be more accurate the more learning repetitions there must be. The learning
algorithm is tested for accuracy using fresh data generated from 25% of the source data. It is possible to see
and plot the covariate points. The correlation is better when train and test data have the highest anticipated
covariate points [24].

4.3. Algorithm pseudocode

The Algorithm 1 defines the method of expectation-maximized value extraction via an iterative
learning process. The procedure begins with a starting value chosen at random from the available data. The
distribution probability is calculated using (1). Any maximal differences in the provided data pattern
cumulatively affect the distributed graph, which has an unimodal structure. The interpretation lambda used for
variance is based on the positive and negative functions.

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 1, March 2024: 201-209
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  205

Algorithm 1. Expectation maximization algorithm

Open
Input Model;
E1=Read_data (Amigos);
Scale_data(N_frames(E1));
Choose random values k=x;
Develop Parm(k);
Update weight;
If(parm(E1) =Exp_Max)
E1=parm_E1;
Else
Update_data;
End loop;
Visualize New_data;
Plot regression;
Close

5. RESULTS AND DISCUSSIONS

5.1. Convergence analysis on physiological signals
Physiological signals such as EEG, ECG, and GSR are analysed for the convergence. Figure 2
analyses ECG, EEG, and GSR data convergence across several iterations. The iterations are done for various
people to analyse the unidentified labels. In Table 1, the calculated results are tabulated for confirmation.

Figure 2. Convergence analysis on ECG, EEG, and GSR

Table 1. Iterations vs. error rate (e) and mean (u)

Iterations Error rate uMin uMax
1 3,548.8753 0.623 2,443.95
2 11,941.6081 3.6651 10,652.47
3 4,149.7602 4.0671 13,505.42
4 2,796.9045 4.2713 15,428.54
5 2,027.8982 4.449 16,822.77
6 942.3074 4.5347 17,470.45
7 372.3683 4.5476 17,726.56
8 126.5081 4.5489 17,813.63
9 0.00 4.5489 17,813.63

5.2. Unique covariate points on the Emo-GEM model

Figure 3, the GEM method assessed the unique covariate points of a single individual under test.
These points are distinct from the rest of the scattered random data [25]. After the given iterations, the error
rate approaches a minimum; the points are extracted. The iterations begin with the maximum error, and the
error rate arranges the u value until the expectation algorithm finds the maximum value.
Affective analysis in machine learning using AMIGOS with Gaussian … (Balamurugan Kaliappan)
206  ISSN: 2089-4864

The entire procedure simplifies extracting unique points from a vast collection. The system can learn
and execute fresh data searches iteratively thanks to the discrepancy between the expected value and the highest
value found. The technique can complete the provided test data in no more than 500 million seconds. The
working model for the first maximization value achieved is formed during the training procedure. Table 1
shows for the specified test sample, the number of iterations ranged from 0.00 after the ninth iteration to
3,548.8753 at the beginning of the random selection of the starting location. It evaluates the statistical data
points about the mean [26].
The Figure 4 shows the graphical representation of population search for a particular test sample over
numerous iterations. The visualization effect is shown in the visualization iterations vs. error rate (e), mean (u)
estimation analysis. Table 2 shows about the obtained emotion label, the overall parameter measurement with
parameter measurements such as are determined.
Figure 5 demonstrates the classification of emotional impact for the given test samples. Anger,
contempt, disgust, happiness, and normal are the four categories into which the test data is divided in the
proposed model. The statistical measurements variance (λ), standard deviation (SD), error rate (e), and mean
(u) aid in classifying the various variables. Table 3 shows the comparison table of existing implementation on
emotional affect detection with proposed GEM algorithm performance and analysis.

Figure 3. ECG, EEG, and GSR unique covariate of single participant

Iterations Error Rate (e) and Mean (u)

9
8
7
6
5
4
3
2
1
0 2,000 4,000 6,000 8,000 10,000 12,000 14,000 16,000 18,000 20,000

Error Rate Iterations

Figure 4. Visualization iterations vs. error rate (e) and mean (u) estimation

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 1, March 2024: 201-209
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  207

Table 2. Emotion labelling with physiological data on different participants

Physio data Emotion label Variance (𝜆) Sigma 1 (SD 𝛴) Sigma 2 (𝛴 Max)
ECG Anger 0.8925 8.7022 134,903.2056
ECG Anger 0.893 9.6402 193,923.7399
EEG Anger 0.879 9.9524 66,890.2988
EEG Anger 0.88375 11.8962 191,742.3546
GSR Anger 0.88875 5.7734 125,302.4221
GSR Anger 0.8845 6.2729 134,375.2957
ECG Contempt 8.2213 0.88825 116,364.1597
EEG Contempt 0.87025 10.8091 130,774.1179
GSR Contempt 0.9 6.4912 115,197.8651
ECG Disgust 0.892 9.141 178,349.1068
ECG Disgust 0.89 9.3346 66,622.7735
EEG Disgust 0.8775 10.8876 171,178.1147
EEG Disgust 0.88525 11.1879 113,150.0861
GSR Disgust 0.89675 7.0367 102,351.8506
GSR Disgust 0.89875 6.6059 179,998.2492
ECG Happy 0.882 9.0123 164,960.485
ECG Happy 0.8905 8.7424 105,189.2817
EEG Happy 0.87375 10.2731 109,640.1184
EEG Happy 0.88125 11.283 164,761.7216
GSR Happy 0.9005 6.7607 158,058.2385
GSR Happy 0.8935 6.777 50,920.1948
ECG Normal 0.8895 10.838 86,024.0168
EEG Normal 0.868 10.6751 145,732.2119
GSR Normal 0.89825 6.4408 129,405.7874

14
Varience(λ), Standard

12
10
Deviation(∑)

8
6
4
2
0
0 5 10 15 20 25 30
Classified Emotions
Lamda (Varience) Sigma 1(SD)

Figure 5. Classification of emotional affect based on test samples

Table 3. GEM algorithm performance and analysis

S. No Input type Ref. Method Categories Statistical measure
1 EEG [5] Neural networks Pleasant, unpleasant, and Mean=0.43,
neutral SD=0.16
2 EDA, HR, [10] KNN, weighted multi- Neutral, cheer, sad, erotic, Mean=0.71,
TEMP dimensional dynamic time and horror SD=0.12
warping (WMD-DTW)
3 ECG, EEG, Proposed GEM Anger, contempt, disgust, Mean=0.60,
GSR happiness, and normal SD=0.80

6. CHALLENGES
The presented work's key problem is dealing with large amounts of data and the processing delay
required for training and testing. The GEM model uses probabilistic distribution and similarity mapping to
determine the relative convergence of grouped data. To scale the data before processing, the system model
should focus on improving the preprocessing stage and feature extraction procedures.

7. CONCLUSION
The GEM algorithm is used to assess emotion identification. The AMIGOS dataset is taken into
consideration for analysis. Analyses are done on physiological signals like the ECG, EEG, and GSR. The
proposed research project focuses on in-depth investigation and the creation of a simple model for emotion
Affective analysis in machine learning using AMIGOS with Gaussian … (Balamurugan Kaliappan)
208  ISSN: 2089-4864

analysis that results in shorter latency. Participants are randomly selected and assessed using data covariance
analysis for ECG, EEG, and GSR along with Emo-GEM and GEM models based on data regression. There are
more distinct correlation points produced to determine emotions the stronger the processing depth-wise
convergence, which leads to data size equality. With 0% error to the maximum iterations, the suggested model
yields a latency of around 438 million second for overall processing. The statistical measurements for detecting
emotions such as anger, contempt, disgust, happiness, and normal are emphasized as mean=0.60, and SD=0.80.
To uncover the detailed variances, the system must also be categorized using a deep learning model.

REFERENCES
[1] J. Han, Z. Zhang, Z. Ren, and B. Schuller, “EmoBed: strengthening monomodal emotion recognition via training with crossmodal
emotion embeddings,” IEEE Transactions on Affective Computing, vol. 12, no. 3, pp. 553–564, Jul. 2021, doi:
10.1109/TAFFC.2019.2928297.
[2] J. Arroyo-Palacios and M. Slater, “Dancing with Physio: a mobile game with physiologically aware virtual humans,” IEEE
Transactions on Affective Computing, vol. 7, no. 4, pp. 326–336, Oct. 2016, doi: 10.1109/TAFFC.2015.2472013.
[3] K. Soltani and R. N. Ainon, “Speech emotion detection based on neural networks,” in 2007 9th International Symposium on Signal
Processing and Its Applications, Feb. 2007, pp. 1–3, doi: 10.1109/ISSPA.2007.4555476.
[4] G. Yang, J. S. Y. Ortoneda, and J. Saniie, “Emotion recognition using deep neural network with vectorized facial features,” in 2018
IEEE International Conference on Electro/Information Technology (EIT), May 2018, pp. 0318–0322, doi:
10.1109/EIT.2018.8500080.
[5] E. Kroupi, J.-M. Vesin, and T. Ebrahimi, “Subject-independent odor pleasantness classification using brain and peripheral signals,”
IEEE Transactions on Affective Computing, vol. 7, no. 4, pp. 422–434, Oct. 2016, doi: 10.1109/TAFFC.2015.2496310.
[6] V. S. Bakkialakshmi and T. Sudalaimuthu, “A survey on affective computing for psychological emotion recognition,” in 2021 5th
International Conference on Electrical, Electronics, Communication, Computer Technologies and Optimization Techniques
(ICEECCOT), Dec. 2021, pp. 480–486, doi: 10.1109/ICEECCOT52851.2021.9707947.
[7] X. Zhang, W. Li, H. Ying, F. Li, S. Tang, and S. Lu, “Emotion detection in online social networks: a multilabel learning approach,”
IEEE Internet of Things Journal, vol. 7, no. 9, pp. 8133–8143, Sep. 2020, doi: 10.1109/JIOT.2020.3004376.
[8] M. Li, L. Xie, Z. Lv, J. Li, and Z. Wang, “Multistep deep system for multimodal emotion detection with invalid data in the internet
of things,” IEEE Access, vol. 8, pp. 187208–187221, 2020, doi: 10.1109/ACCESS.2020.3029288.
[9] M.-H. Hoang, S.-H. Kim, H.-J. Yang, and G.-S. Lee, “Context-aware emotion recognition based on visual relationship detection,”
IEEE Access, vol. 9, pp. 90465–90474, 2021, doi: 10.1109/ACCESS.2021.3091169.
[10] M. R. Islam et al., “Emotion recognition from EEG signal focusing on deep learning and shallow learning techniques,” IEEE Access,
vol. 9, pp. 94601–94624, 2021, doi: 10.1109/ACCESS.2021.3091487.
[11] A. Albraikan, D. P. Tobon, and A. El Saddik, “Toward user-independent emotion recognition using physiological signals,” IEEE
Sensors Journal, vol. 19, no. 19, pp. 8402–8412, Oct. 2019, doi: 10.1109/JSEN.2018.2867221.
[12] R. Qayyum et al., “Android based emotion detection using convolutions neural networks,” in 2021 International Conference on
Computational Intelligence and Knowledge Economy (ICCIKE), 2021, pp. 360–365, doi: 10.1109/ICCIKE51210.2021.9410768.
[13] V. S. Bakkialakshmi and T. Sudalaimuthu, “Emo-Net artificial neural network: a robust affective computing prediction system for
emotional psychology using AMIGOS,” Indian Journal of Computer Science and Engineering, vol. 13, no. 4, pp. 1040–1055, Aug.
2022, doi: 10.21817/indjcse/2022/v13i4/221304034.
[14] P. Sarkar and A. Etemad, “Self-supervised ECG representation learning for emotion recognition,” IEEE Transactions on Affective
Computing, vol. 13, no. 3, pp. 1541–1554, Jul. 2022, doi: 10.1109/TAFFC.2020.3014842.
[15] L. Song and W. Luo, “Self-supervised learning of visual odometry,” in 2020 International Conference on Information Science,
Parallel and Distributed Systems (ISPDS), Aug. 2020, pp. 5–9, doi: 10.1109/ISPDS51347.2020.00009.
[16] V. S. Bakkialakshm and T. Sudalaimuthu, “Dynamic cat-boost enabled keystroke analysis for user stress level detection,” in 2022
International Conference on Computational Intelligence and Sustainable Engineering Solutions (CISES), May 2022, pp. 556–560,
doi: 10.1109/CISES54857.2022.9844331.
[17] H. Ranganathan, S. Chakraborty, and S. Panchanathan, “Multimodal emotion recognition using deep learning architectures,” in
2016 IEEE Winter Conference on Applications of Computer Vision (WACV), Mar. 2016, pp. 1–9, doi:
10.1109/WACV.2016.7477679.
[18] S. Mishra, S. Kumar, S. Gautam, and J. Kour, “Real time expression detection of multiple faces using deep learning,” in 2021
International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE), Mar. 2021, pp. 537–
542, doi: 10.1109/ICACITE51222.2021.9404561.
[19] H. Wan, H. Wang, B. Scotney, and J. Liu, “A novel gaussian mixture model for classification,” in 2019 IEEE International
Conference on Systems, Man and Cybernetics (SMC), Oct. 2019, pp. 3298–3303, doi: 10.1109/SMC.2019.8914215.
[20] P. Barros, N. Churamani, E. Lakomkin, H. Siqueira, A. Sutherland, and S. Wermter, “The OMG-emotion behavior dataset,” in 2018
International Joint Conference on Neural Networks (IJCNN), Jul. 2018, pp. 1–7, doi: 10.1109/IJCNN.2018.8489099.
[21] V. S. Bakkialakshmi, T. Sudalaimuthu, and B. Umamaheswari, “Emo-Spots: detection and analysis of emotional attributes through
bio-inspired facial landmarks,” in International Conference on IoT, Intelligent Computing and Security. Lecture Notes in Electrical
Engineering, Singapore: Springer, 2023, pp. 103–115, doi: 10.1007/978-981-19-8136-4_9.
[22] W. M. B. Henia and Z. Lachiri, “Emotion classification in arousal-valence dimension using discrete affective keywords tagging,”
in 2017 International Conference on Engineering and MIS (ICEMIS), May 2017, pp. 1–6, doi: 10.1109/ICEMIS.2017.8272991.
[23] V. S. Bakkialakshmi, T. Sudalaimuthu, and S. Winkler, “Effective prediction system for affective computing on emotional
psychology with artificial neural network,” EasyChair Preprint. 2022.
[24] Y. Baveye, C. Chamaret, E. Dellandrea, and L. Chen, “Affective video content analysis: a multidisciplinary insight,” IEEE
Transactions on Affective Computing, vol. 9, no. 4, pp. 396–409, Oct. 2018, doi: 10.1109/TAFFC.2017.2661284.
[25] D. Nikolova, P. Georgieva, P. Petkova, and A. Manolova, “ECG-based emotion recognition: overview of methods and applications,”
in ANNA 2018 - Advances in Neural Networks and Applications 2018, 2018, pp. 118–122.
[26] K. Maehara and K. Fujinami, “Psychological effects on positional relationships between a person and a human-following robot,”
in 2018 IEEE 24th International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA), Aug.
2018, pp. 242–243, doi: 10.1109/RTCSA.2018.00041.

Int J Reconfigurable & Embedded Syst, Vol. 13, No. 1, March 2024: 201-209
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  209

BIOGRAPHIES OF AUTHORS

Balamurugan Kaliappan is a research scholar at Hindustan University (Hindustan

Institute of Technology and Science). He did his M.BA. postgraduate and is a industry
professional and an experienced cyber security expert. He works as senior delivery manager
position in L&T technologies Private Limited. He can be contacted at email:
[email protected].

Bakkialakshmi Vaithialingam Sudalaiyadumperumal received her doctorate in

computer science engineering from Hindustan Institute of Technology and Science (deemed to
be Hindustan University), India, and her master of science degree in computer science from
Anna Adarsh College for Women of India in 2004 and received the master of philosophy degree
in computer science from the Bharadhidasan University of India in 2005. She received a master
of engineering in computer science engineering from Anna University Affiliation College, India,
in 2007. She is an assistant professor at the SRM Institute of Science and Technology,
Kattankulathur, India. She is an international research traveler who traveled to (Austria) Europe
for research collaboration and was a resource in many seminars and workshops. Her research
area is affective computing on emotional psychology. She can be contacted at email:
[email protected].

Sudalaimuthu Thalavaipillai is working as a professor in the School of Computing

Science, Hindustan Institute of Technology and Science, Chennai, India. He completed his Ph.D.
from the Hindustan Institute of Technology and Science, Chennai, India. He is a certified ethical
hacker. He has published 50 research articles in reputed international journals and conferences.
He is granted both Indian and Australian patents. He obtained many awards in his career,
including the Pearson Award in the best teacher category and the Top Innovator Award for patent
rights. His research areas include cyber network security, grid and cloud computing, and
machine learning. He is a lifetime member of CSI, ACM, and IEEE. He can be contacted at
email: [email protected].

Affective analysis in machine learning using AMIGOS with Gaussian … (Balamurugan Kaliappan)