IRJSTEM-Online Class Engagement Detection Model

You might also like

Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 12

Volume X, No.

X | Month Year

Online Class Engagement Detection Model


Pongsathorn Cherdsom1, Wanida Kanarkard2
Khon Kaen University, Khon Kaen Thailand
pongthorn@kkumail.com1, wanida@kku.ac.th2

ABSTRACT ARTICLE INFO


The pandemic of the coronavirus (COVID-19) has affected the Received :
education system, disrupting all from traditional classrooms to online Revised :
classrooms. This complicates tracking involvement in online classes Accepted :
even further. A student's separation from education is the most serious
scenario that can occur, in addition to damaging the efficiency of the KEYWORDS
learners. Teachers should indeed be regularly informed about student Engagement detection,
participation so that they may customize their instruction to the online Emotion recognition, Class
learning environment. The construction of a model to monitor online engagement
classroom participation using neural networks is presented in this paper.
Learners' faces are used by Convolutional Neural Networks (CNN),
which uses a base model from the Keras website, with three levels of
engagement: disengagement, normal engagement, and high engagement.
The results showed that using Optimizer Ranger in combination with
Fully connected FC: 50-100, accuracy is increased by 16.51 and loss is
0.31 compared to baseline model.

Suggested Citation (APA Style 7th Edition):

https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

INTRODUCTION
The coronavirus disease 2019 (COVID-19) outbreak began in Wuhan, China (World Health Organization, 2020)
in December 2019 and has since expanded to over 200 nations (worldometers, 2020) across all sectors, including
education. From a traditional classroom to an online classroom This makes it harder to track engaged behavior.
Studying in a virtual classroom is far more challenging than studying in a traditional classroom.
Different methodologies of analyzing and tracking student participation were classified into first, automated
approaches: employing tools or devices to evaluate and track participation from (Dewan et al., 2002). To aid in
analysis and assessment, such as through the use of sensor data. Website access data, picture data, or films filmed in
class utilizing facial and body motions Second, semi-automatic eye motions, which is a classroom Q&A session.
third, the most often used human-operated evaluation for determining learning and involvement levels. by relying
on instructor evaluations or student evaluations Even if human action is frequently employed in the typical
classroom, adjustments in teaching approaches are suitable in exceptional scenarios. As a result, technology and
automatic interaction tracking are becoming more prevalent.
Artificial Intelligence (AI) is a branch of computer science that is now being utilized to monitor and measure
student involvement. Machine learning and deep learning skills are used to assess engagement by using extremely
accurate and dependable picture or video processing technology known as computer vision. of pupils from
(Murshed et al., 2002) well-known algorithms Convolutional Neural Network (CNN) is a type of neural network
that has the capacity to extract features from complicated pictures and is suited for unstructured data. One model
and three more models were made from other publications. To examine learners' involvement, a model was
utilized, and the model's outputs were compared to the learners' faces. Also, Disengagement, Normal Engagement,
and High Engagement were the three different levels of involvement. The created model was more accurate than
the three models taken from the other papers, with an accuracy of 92.33 percent.
TensorFlow Torch, Keras, and SciKit-Learn are just a few of the fascinating frameworks and packages available
today for model construction. The Kears library, a TensorFlow-based API, was chosen for model building in this
post because of its simplicity, flexibility, and power. To submit a concept, experiment, or do a study. It is utilized
by CERN, NASA, the National Institutes of Health, and many other scientific institutions across the world (Keras,
2020).
To create efficient models The parameters (Hyperparameters) must be optimized, and there are several
interesting parameters to alter for the model to work. Learn about the models that have been developed as well as
some of the most popular ready-made models for visualizing metastatic cancer. (Zhonget al., 2020) were used to
classify cancer photos, and (Igiri al., 2021) compared outcomes using machine learning rates ranging from 0.10 to
0.80. The following parameter, Optimizer (Eustace al., 2018), was introduced. SGD vanilla, SGD momentum +
nesterov, SGD momentum, AdaGrad, RMSProp, Adam, AdaDelta, Adamax, Nadam, SGD vanilla, SGD
momentum + nesterov, SGD momentum, AdaGrad, RMSProp, Adam, AdaDelta, Adamax, Nadam (Zheng al.,
2021) DenseNet, Resnet, and VGG16 were used to assess the implementation of optimization methods such as
RAdam in classification. of the pictures labeled as malignant or not (Melinte al., 2020), Optimizer use Rectified
Adam has been shown to predict human emotions, and (Suresha and Parthasarathy, 2020) showed a model for
identifying Alzheimer's disease that incorporates an optimizer similar to Rectified Adam, as well as the next
parameter, Batch Size (Kandel al., 2020), to compare the results of the Batch Sizes 16,32,64,128,256. According to
CNN.
The design of a model for analyzing learner engagement behavior is presented in this study. Using a CNN
algorithm can algorithms to optimize learning rate, batch size, and optimizer parameters. Customization in the
feature isolation layer and the fully linked layer is also possible. and categorized the degrees of engagement into
three categories: Disengagement, Normal Engagement, and High Engagement are the three levels of engagement.
The dataset that was used to train the AffectNet model is the most commonly utilized and powerful dataset in this
field.

2
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

OBJECTIVES
Develop a model and evaluate the performance for analyzing and tracking student participation.
LITERATURE REVIEW
Even now, Quiet was categorized as disengagement, Calm, Relaxed, Bored, Depressed, Tired was
categorized as low engagement, Happy, Content, Delighted, Joyful, Satisfied, was classed as normal
engagement, Angry, Ashamed, Fearful, Nervous, Enthusiastic, Excited was classified as high engagement,
and Surprise, involvement was classified as strong engagement. (Dewan al., 2018) categorized by boredom,
confusion, impatience, described as Disengagement, and (Dash al., 2019) linked to behavior by conduct that
displays interest, as illustrated in Figure 1. Engagement is defined as a feeling of fulfillment or bewilderment.
Emotional contentment and a lack of behavioral interest Disengagement is the result of confusion and
boredom.

Figure 1. The link between emotion and engagement (Khawlah al., 2021).
Emotions were grouped into 2-5 levels to adhere to learner involvement analysis, with 2 levels of
Engagement and Disengagement (Dewan al., 2018) (Saswat al., 2019) (Sudhir al., 2018) (Nezami al., 2018)
(Zhang al., 2019) (Chang al., 2018) (Huang al., 2019) (Ömer al., 2021), and 3 degrees of Disengagement.
Normal and high levels of involvement (Dewan al., 2018) (Murshed al., 2019) (Huang al., 2019) (Sümer al.,
2021) (Sharma al., 2019) (Booth al., 2017) (Liu al., 2018) (Ashwin al., 2020). There are four stages, each
with a very low degree of interaction. typical interaction Disengagement is one of five degrees of
involvement, which includes high engagement (Chang al., 2018) (Kaur al., 2018) (Yan-Ying al., 2019)
(Michael al., 2021) (Chang al., 2018) (Liao al., 2021) (Gupta al., 2016) (Alnafjan al., 2017) (Abedi al., 2021)
(Jiajia al., 2016). There is less involvement. High engagement Normal engagement Highest level of
involvement (Khawlah al., 2021)
Convolutional neural network (CNN) is a special type of neural network. Feature extraction from
complicated photos is ideal for unstructured data. As demonstrated in Figure 2, the CNN design is made up
of three layers. 1) CNN's Convolution layer (Convolutional Layer) is a crucial layer. The Cornell Filter
(Kernel Filter) and Cornell Filter (Kernel Filter) are the layers that bring out the features of the data (Kernel
size). with sizes ranging from 3X3 to 5X5, 7X7, and so on. 2) The Pooling Layer, which includes max
pooling, min pooling, sum pooling, and average pooling, extracts relevant sections of the data, and decreases
its size (Average Pooling). The data from the preceding two-dimensional layer is transformed into one-
dimensional data by the Fully Connected Layer which before each and every node is linked.

3
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Figure 2. Structure of CNN (Sengupta al., 2020).

Demonstrated mood prediction modeling utilizing the AffectNet dataset and the AlexNet model as a foundation for
future model development. Figure 3 depicts the structure presented in this study. AlexNet had a precision of 58.00 percent,
while the subsequently evolved model had a precision of 93.80 percent, according to the findings of the trial (Jalal al., 2019).

Figure 3. The structure of the model presented (Jalal al., 2019).

Conducted a study and research into learners' engagement in e-learning through the use of an algorithm that
is based on 1) All Convolutional Network (All-CNN), 2) Network in Network (NiN-CNN), 3) Very Deep
Convolutional (VD-CNN), and 4) are developed regions are the four types of convolutional neural network (CNN).
More findings from this study Figure 4 depicts the structure and parameters of each model utilized in the
investigation. Participation was categorized into three levels in the study: 1) There is no involvement. (not-engaged)
2) Consistent involvement (normally engaged) 3) Extremely involved (involved) and, in this study, utilized to
examine The analysis in this study employed datasets to construct them based on the students' videos taken via
webcams. DAiSEE is the model used in the analysis. The investigation found that All-CNN has a 75.97 percent
accuracy, NiN-CNN has an 83.22 percent accuracy, VD-CNN has an 86.45 percent accuracy, and 4) are developed
portions. The accuracy was 92.33 percent in addition to this research (Murshed al., 2019).

Figure 4. Structure and parameters of each model (Murshed al., 2019).

4
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Learner involvement in online classrooms has been explored and researched. An algorithm was employed
to examine the participants' participation in this study. It is based on the following five CNN models: 1) ConvAll
2) The VDConv 3) NiNConv NiNConv NiNConv NiNConv Ni 4) Convolutional Con-Pool (CPConv)
5) Convolutional Cluster Munitions (CMConv). The dataset was utilized to develop a model, DAiSEE, consisting
of a dataset of 80 males, 32 women, and a total of 9068 movies. The model parameters used in the study are
displayed in Fig. 5. The movie is around 10 minutes long, with a resolution of 1920 × 1080 pixels and a frame rate
of 30 frames per second. The study divided involvement into two categories: disengaged and engaged. The
following are the findings of the investigation: The accuracy of the AllConv model was 81.80 percent, the VDConv
model was 79.20 percent, the NiNConv model was 81.20 percent, the CPConv model was 83.60 percent, and the
CMConv model was 91.10 percent (Saswat al., 2019).

Figure 5. Model structure and parameters (Saswat al., 2019).

EXPERIMENTS
TOOLS
The Keras and Tensorflow libraries are used in this study to create a Python-based model with an Intel Core i5-
10400F CPU, 32GB of main memory, Nvidia RTX 3060 (12GB) graphics processor, and Ubuntu as the operating
system. LTS Version 20.04

DATASET
In this study, the AffectNet dataset was used to learn the model. The images were 224 x 224 x 3 in size, with a
total of 959,907 images, 11 labeled 8 emotions: neutral, happiness, sadness, excitement, fear, disgust, anger,
contempt, and No face, no emotion, not sure (Mollahosseini al., 2017) as shown in Fig. 6, and were grouped in
accordance with the research into 3 groups at the level of engagement: Disengagement, Normal engagement, High
engagement

5
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Figure 6. Image of a face from the AffectNet dataset. (A) neutral 


(B) happiness (C) sadness (D) excitement 
(E) fear (F) disgust (G) anger (H) contempt (Mollahosseini al., 2017).

DATA PRE-PROCESSING
Regrouping from 8 emotions into 3 levels of participation is shown in Table 1. Regrouping for research use.
Select a total of 96,111 images for each mood level. which is the smallest of the three levels to equalize all three
levels and uses K-fold cross-validation. to share information for learning and mock tests by dividing the data into 5
groups (K=5), 3 sets, ratio 75:15:10, respectively (Keras al., 2020) consisted of a 75% training set with 216,149
images, a validation set 15% of 43,251 images, and a test data set (test set) 10% of 28,833 images, as shown in
Table 2. grouping of data for research use.

Table 1
Grouping of data for research use.
New label Label from dataset Amount Total
contempt 4,252
Disengagement 222,768
neutral 218,516
happiness 381,150
Normal engagement sadness 46,813 433,156
boredom 5,193
excitement 32,052
High engagement 96,111
anger 53,382

Table 2
Data Split for model learning.
Class Training Set (75%) Validation Set (15%) Test Set (10%)
Disengagement 72,083 14,417 9,611

Normal engagement 72,083 14,417 9,611


High engagement 72,083 14,417 9,611
Total 216,149 43,251 28,833

6
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

EVALUATION
The efficacy of model learning was measured using Accuracy. accuracy between what is predicted and
what is happening the calculation equation is as follows.

TP+TN
Accuracy=
TP+TN + FP+ FN

TP is the value that the forecast is positively correct.


TN is a negative accurate forecast value.
FP is a positive predictive value.
FN is the predictive error value.

EXPERIMENT SETUP
After grouping the data use the function ImageDataGenerator to one-hot encode and convert data to Numpy and
use flow_from_directory Fetch data from all three groups, set input_size to 224 x 224, set categorical class_mode,
and set batch_size for each cycle as specified, set 15 epochs every trial, and use the K-fold cross validation
technique: K=5 to divide the 5 data sets to test the validity of the model. As shown in Figure 7.

Figure 7. model learning process.

RESULTS
Developing a model to monitor participation in online classes. has adopted a structure Modeled
from the Keras website is the baseline model structure for model development. The experiments are divided
into 3 groups: The first group is to adjust the structure of the feature extraction layer of the image. The
second group consisted of experiments with individual parameters: Learning Rate, Batch Size, Optimizer,
Fully connected. The third group was a two-parameter matching test, where groups 1-3 were tested
sequentially. and the best-performing model from the previous test was applied to the next trial.
Group 1 is to adjust the structure of the feature extraction layer of the image. As shown in Table 3,
the results of the experimental manipulation of the Featured extraction layer. The experimental patterns were
divided into six types: 1) a steady increase in the number of Neural Network (NN), 2) a consistent increase
and a decrease in the number of NN equal to an increase, and 3) a constant decrease in the number of NN 4)
The number of NNs is the same 5) Number of NN alternate (two values) 6) The number of NN increasing
and decreasing are unequal as shown in Figure 8.
7
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

The Feature extraction column shows the number of Convolution2D filters. Each layer consists of
Convolution2D with different number of Filters, kernel size 3x3, Activation Function Relu, and
MaxPooling2D Pool Size 2x2 the same. The best results were Feature extraction Filter: 128-256-512
Accuracy 79.62 and Loss 0.52 which worked. Accuracy is close to Filter: 128-256-512-1024 but Loss is
higher.
Table 3
The result of the feature extraction layer optimization.
Learnin
Optimize Batch Accuracy Loss
Feature extraction Fully connected g
r size Validation Validation
rate
(1) 64-64-64-64 FC: 0 (output 3) adam 80 0.0001 70.70 0.66
(2) 32-16-32-16 FC: 0 (output 3) adam 80 0.0001 70.21 0.67
(3) 1024-512-256 FC: 0 (output 3) adam 80 0.0001 73.05 0.63
(4) 128-256-512-1024 FC: 0 (output 3) adam 80 0.0001 78.98 0.87
(5) 128-256-512 FC: 0 (output 3) adam 80 0.0001 79.62 0.52

Figure 8. The pattern of increasing/decreasing the number of NN in the experiment


(1) a steady increase in the number of Neural Network (NN)
(2) a consistent increase and a decrease in the number of NN equal to an increase, and
(3) a constant decrease in the number of NN (4) The number of NNs is the same
s(5) Number of NN alternate (two values) (6) The number of NN increasing and decreasing are unequal.

Group 2 Parameter experiment


Learning Rate
The results of the experiments are shown in Table 4, showing the results of the learning rate
parameters. The learning rate values were 0.10,0.15,0.20, 0.25,0.30,0.35,0.01. The experimental results
were Accuracy 33.33 and Loss 1.10 could not be dropped from 33.33, Learning Rate 0.0001 gave the best
result. The result was 80.75 accuracy and 0.51 loss and was the base model used.

Table 4
The results of the experimental parameters of Learning Rate.
Batch Learning Accuracy Loss
Fully connected Optimizer
size rate Validation Validation
FC: 0 (output 3) adam 80 0.01 33.33 1.10
FC: 0 (output 3) adam 80 0.001 79.78 0.53
FC: 0 (output 3) adam 80 0.0001 80.75 0.51
FC: 0 (output 3) adam 80 0.00002 79.75 0.51

8
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Batch Size
The experimental results are shown in Table 5, showing the experimental results of the Batch Size
parameters, which are as follows: 16, 32, 64, 80, 128, 256 The values that gave the best results were Batch
Size 128, the results were Accuracy 81.27 and Loss 0.50

Table 5
The results of the experimental parameters of Batch Size.
learning Accuracy Loss
Fully connected optimizer batch size
rate Validation Validation
FC: 0 (output 3) adam 16 0.0001 80.22 0.62
FC: 0 (output 3) adam 32 0.0001 80.39 0.56
FC: 0 (output 3) adam 64 0.0001 80.42 0.53
FC: 0 (output 3) adam 80 0.0001 81.00 0.51
FC: 0 (output 3) adam 128 0.0001 81.27 0.50
FC: 0 (output 3) adam 256 0.0001 80.78 0.49

Optimizer
The experimental results are shown in Table 6. The Optimizer parameters were Adam and
Adamax, Nadam, DiffGrad, Ranger, Rectified Adam, all of which were developed after Adam. The best
experimental result was Ranger. The experimental result was Accuracy 81.04 and Loss 0.49

Table 6
The results of the experimental parameters of Optimizer.
Learning Accuracy Loss
Fully connected Optimizer Batch size
rate Validation Validation
FC: 0 (output 3) Adam 80 0.0001 80.97 0.51
FC: 0 (output 3) Adamax 80 0.0001 80.69 0.49
FC: 0 (output 3) Nadam 80 0.0001 80.90 0.52
FC: 0 (output 3) DiffGrad 80 0.0001 80.67 0.49
FC: 0 (output 3) Ranger 80 0.0001 81.04 0.49
FC: 0 (output 3) Rectified Adam 80 0.0001 80.85 0.52

Fully connected
The results are shown in Table 7, showing the results of the fully connected parameter,
the experimental patterns were divided into six types: 1) a steady increase in the number of Neural Network
(NN), 2) a consistent increase and a decrease in the number of NN equal to an increase, and 3) a constant
decrease in the number of NN 4) The number of NN is the same 5) Number of NN alternate (two values)
6) The number of NN increasing and decreasing are unequal as shown in Figure 8. which is the number of
neural networks, the best results were FC: 50-100 and FC: 50-50. The results were Accuracy 82.20 and
Loss 0.49, and when comparing the results in each epoch, FC: 50-50 had more overfit than FC: 50-100.

Table 7
The results of the experimental parameters of Fully connected.
Learning
Fully connected Optimizer Batch size Accuracy Validation Loss Validation
rate
FC: 32 adam 80 0.0001 81.07 0.50
FC: 64-32 adam 80 0.0001 81.20 0.51
FC: 64-64 adam 80 0.0001 78.54 0.61
FC: 50-100 adam 80 0.0001 82.20 0.49
FC: 50-50 adam 80 0.0001 82.20 0.49
9
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

FC: 50-100-50 adam 80 0.0001 82.12 0.50


Group 3 a two-parameter matching test
 Optimizer + Batch Size (A)
 Optimizer + Learning Rate (B)
 Batch Size + Learning Rate (C)
 Optimizer + Fully connected (D)
 Batch Size + Fully connected (E)
 Learning Rate + Fully connected (F)
The results are shown in Table 8, matching the results of the third experimental group. The best
performing parameter in the second group in each experiment was matched to two parameters. The best is
Optimizer + Fully Connected. The result is 82.30 accuracy and 0.46 loss.

Table 8
Results of the third group of experiments by pairing 2 parameters.
Loss
Batch Learning Accuracy
Lab Fully connected Optimizer Validatio
size rate Validation
n
A FC: 0 (output 3) Ranger 128 0.0001 80.66 0.49
B FC: 0 (output 3) Ranger 80 0.0001 80.89 0.48
C FC: 0 (output 3) adam 128 0.0001 80.91 0.50
D FC: 50-100 (output 3) Ranger 80 0.0001 82.30 0.46
E FC: 50-100 (output 3) adam 128 0.0001 82.16 0.48
F FC: 50-100 (output 3) adam 80 0.0001 81.72 0.50

Comparison of the best results from the experimental group 1 - 3 with the baseline model.
As shown in Table 9, it also shows the increase/decrease in Accuracy and Loss amount.
Experiment 1 belongs to the baseline model.
Experiment 2 This was the result with the best results in the trial group 1.
Experiment 3 - 6 This was the experiment with the best results in the second trial group.
Experiment 7 - 18 This was the result of the experiment in the third experimental group.

CONCLUSIONS
This study examines the creation of a model for tracking online class participation. by separating
participation into three levels using convolutional neural networks (Convolutional Neural Networks): There will be
no participation. involvement at the customary level involvement at a high level The AffectNet dataset was utilized
to train the model. The Training set, Validation set, and Test set are the proportions of the 75:15:10 division,
respectively. The experiments were separated into three groups: the feature extraction layer structure optimization,
the individual parameter trials, and the feature extraction layer structure optimization. Completely interconnected
Group 3 was an experiment in which two factors were paired and the trials were carried out in order. Groups 1-3
are the experimental groups. It's also the best-performing model from the prior round of experiments. Let's continue
with the experiment in the next group. The findings of the trial revealed that experimental groups 1-3 produced
superior outcomes. The model's performance, Accuracy of 82.30, a 16.51 improvement from the baseline model,
and Loss of 0.46, a drop from the baseline model, yielded the best results when compared to the basis trial. begin
with 0.31

10
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Data Augmentation may be utilized in the next experiment to boost the quantity of data used in the model's
learning. This might lead to improved model performance and the adoption of additional premade models in the
future.

REFERENCES
World Health Organization: WHO. (2020). Q&A on coronaviruses (COVID-19).
https://www.who.int/emergencies/diseases/novel-coronavirus-2019/COVID-19 CORONAVIRUS
PANDEMIC. (2020). https://www.worldometers.info/coronavirus/
Keras. (2020). https://keras.io/
Keras. (2020). Simple MNIST convnet. https://keras.io/examples/vision/mnist_convnet/
M. A. A. Dewan, F. Lin, D. Wen, M. Murshed and Z. Uddin, (2018). A Deep Learning Approach to Detecting
Engagement of Online Learners, IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced &
Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of
People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), 1895-1902.
Murshed, M., Dewan, M. A. A., Lin, F., & Wen, D. (2019). Engagement Detection in e-Learning Environments
using Convolutional Neural Networks. IEEE Intl Conf on Dependable, Autonomic and Secure Computing,
Intl Conf on Pervasive Intelligence and Computing, Intl Conf on Cloud and Big Data Computing, Intl Conf
on Cyber Science and Technology Congress (DASC/PiCom/CBDCom/CyberSciTech), 80–86.
Qiu, G., Yu, X., Sun, B., Wang, Y., & Zhang, L. (2020). Metastatic Cancer Image Classification Based On Deep
Learning Method.
Zhong, Z., Zheng, M., Mai, H., Zhao, J., & Liu, X. (2020). Cancer image classification based on DenseNet model.
Igiri, Chinwe & Uzoma, Anyama & Ita, Silas. (2021). Effect of Learning Rate on Artificial Neural Network in
Machine Learning. International Journal of Engineering Research.
E. M. Dogo, O. J. Afolabi, N. I. Nwulu, B. Twala and C. O. Aigbavboa, (2018). A Comparative Analysis of
Gradient Descent-Based Optimization Algorithms on Convolutional Neural Networks, International
Conference on Computational Techniques, Electronics and Mechanical Systems (CTEMS), 92-99.
Z. Wan, Z. Yuxiang, X. Gong, Zhanghuali and B. Yu, (2021). DenseNet model with RAdam optimization
algorithm for cancer image classification, IEEE International Conference on Consumer Electronics and
Computer Engineering (ICCECE), 771-775.
Melinte, Octavian & Vladareanu, Luige. (2020). Facial Expressions Recognition for Human-Robot Interaction
Using Deep Convolutional Neural Networks with Rectified Adam Optimizer.
H. S. Suresha and S. S. Parthasarathy, (2020). Alzheimer Disease Detection Based on Deep Neural Network with
Rectified Adam Optimization Technique using MRI Analysis, Third International Conference on Advances
in Electronics, Computers and Communications (ICAECC), 1-6.
Kandel, Ibrahem & Castelli, Mauro. (2020). The effect of batch size on the generalizability of the convolutional
neural networks on a histopathology dataset. ICT Express.
Altuwairqi, K., Jarraya, S. K., Allinjawi, A., & Hammami, M. (2021). A new emotion–based affective model to
detect student’s engagement.
Dewan, M. A. A., Lin, F., Wen, D., Murshed, M., & Uddin, Z. (2018). A Deep Learning Approach to Detecting
Engagement of Online Learners.
Dash, S., Akber Dewan, M. A., Murshed, M., Lin, F., Abdullah-Al-Wadud, M., & Das, A. (2019). A Two-Stage
Algorithm for Engagement Detection in Online Learning. International Conference on Sustainable
Technologies for Industry 4.0 (STI), 1–4.
Mane, S. S., & Surve, A. R. (2018). Engagement Detection using Video-based Estimation of Head Movement.
Nezami, O. M., Dras, M., Hamey, L., Richards, D., Wan, S., & Paris, C. (2018). Automatic Recognition of Student
Engagement using Deep Learning and Facial Expression.
Zhang, Z., Li, Z., Liu, H., Cao, T., & Liu, S. (2019). Data-driven Online Learning Engagement Detection via Facial
Expression and Mouse Behavior Recognition Technology.
Chang, C., Zhang, C., Chen, L., & Liu, Y. (2018). An Ensemble Model Using Face and Body Tracking for
Engagement Detection.
Huang, T., Mei, Y., Zhang, H., Liu, S., & Yang, H. (2019). Fine-grained Engagement Recognition in Online
11
https://irjstem.com
International Research Journal of Science, Technology, Education, and Management
Volume X, No. X | Month Year

Learning Environment.
Sümer, Ö., Goldberg, P., D’Mello, S., Gerjets, P., Trautwein, U., & Kasneci, E. (2021). Multimodal Engagement
Analysis from Facial Videos in the Classroom.
Sharma, P., Joshi, S., Gautam, S., Maharjan, S., Filipe, V., & Reis, M. J. C. S. (2019). Student Engagement
Detection Using Emotion Analysis, Eye Tracking and Head Movement with Machine Learning.
Booth, B. M., Ali, A. M., Narayanan, S. S., Bennett, I., & Farag, A. A. (2017). Toward active and unobtrusive
engagement assessment of distance learners.
Liu, Y., Chen, J., Zhang, M., & Rao, C. (2018). Student engagement study based on multi-cue detection and
recognition in an intelligent learning environment. Multimedia Tools and Applications
T. S., A., & Guddeti, R. M. R. (2020). Automatic detection of students’ affective states in classroom environment
using hybrid convolutional neural networks.
Kaur, A., Mustafa, A., Mehta, L., & Dhall, A. (2018). Prediction and Localization of Student Engagement in the
Wild.
Li, Y.-Y., & Hung, Y.-P. (2019). Feature Fusion of Face and Body for Engagement Intensity Detection.
Thiruthuvanathan, M., Krishnan, B., & Rangaswamy, M. (2021). Engagement Detection through Facial Emotional
Recognition Using a Shallow Residual Convolutional Neural Networks.
Chang, C., Zhang, C., Chen, L., & Liu, Y. (2018). An Ensemble Model Using Face and Body Tracking for
Engagement Detection.
Liao, J., Liang, Y., & Pan, J. (2021). Deep facial spatiotemporal network for engagement prediction in online
learning.
Gupta, A., Jaiswal, R., Adhikari, S., & Balasubramanian, V. (2016). DAISEE: Dataset for Affective States in E-
Learning Environments.
Alnafjan, A., Hosny, M., Al-Wabil, A., & Al-Ohali, Y. (2017). Classification of Human Emotions from
Electroencephalogram (EEG) Signal using Deep Neural Network.
Abedi, A., & Khan, S. S. (2021). Improving state-of-the-art in Detecting Student Engagement with Resnet and TCN
Hybrid Network.
Li, J., Ngai, G., Leong, H. V., & Chan, S. C. F. (2016). Multimodal Human Attention Detection for Reading from
Facial Expression, Eye Gaze, and Mouse Dynamics.
Sengupta, S., Basak, S., Saikia, P., Paul, S., Tsalavoutis, V., Atiah, F., Ravi, V., & Peters, A. (2020). A review of
deep learning with special emphasis on architectures, applications and recent trends.
Jalal, M. A., Mihaylova, L., & Moore, R. K. (2019). An End-to-End Deep Neural Network for Facial Emotion
Classification. International Conference on Information Fusion (FUSION), 1–7.
Mollahosseini, A., Hasani, B., & Mahoor, M. H. (2017). AffectNet: A Database for Facial Expression, Valence,
and Arousal Computing in the Wild.

12
https://irjstem.com

You might also like