INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 865
www.rsisinternational.org
Development of Face Expression Recognition Model to Support
Learning Feedback in Higher Education
Muhammad Firdaus Mustapha
1*
, Siti Haslini Ab Hamid
2
1
Faculty of Computer and Mathematical Sciences, Universiti Teknologi Mara Cawangan Kelantan,
Bukit Ilmu, 18500 Machang, Kelantan, Malaysia
2
Department of Information Technology, FH Training Center, 16800 Pasir Puteh, Kelantan Malaysia
*Corresponding Author
DOI: https://dx.doi.org/10.47772/IJRISS.2025.91100070
Received: 14 November 2025; Accepted: 20 November 2025; Published: 29 November 2025
ABSTRACT
Face expressions offer a non-verbal channel for understanding student engagement and feedback in higher
education learning environment. With the rise of affective computing, face expression recognition (FER)
applications have gained attention for their ability to the recognize and respond to learners’ emotional cues in
real time. Nevertheless, developing a stable FER model often involves complex deep learning architectures and
large-scale annotated datasets. Therefore, this study presents the development of a FER model using Google
Teachable Machine (GTM) to support learning feedback in higher education. The proposed FER model can
classify five categories of face expressions. A dataset comprising 600 face images was collected and divided into
85% for training and 15% for validation/testing. Model performance was evaluated using accuracy, precision,
recall and F1-score metrics. The confusion matrix showed reliable performance for all face expression categories,
validating the effectiveness of GTM for accessible FER model.
Keywords: Face Expression Recognition, Google Teachable Machine, Higher Education, Learning Feedback
INTRODUCTION
The emotional state of learners plays a crucial role in their ability to acquire, retain, and apply knowledge.
Emotions such as frustration, boredom, interest, or confusion can impact concentration and motivation in
significant ways [1]. Face expressions, as a direct and observable indicator of emotion, offer a non-verbal channel
through which student feedback can be interpreted in real time [2]. Face expression is one of the most informative
non-verbal cues in human communication and emotional recognition [3]. Mehrabian [4] claimed that 93% of the
emotional meaning is transmitted as follows: 55% come from facial expression, 38% come from vocal expression
and 7% come from verbal expression.
In the context of higher education, monitoring the evolving face expressions of students over time is crucial for
gaining insights into their engagement, emotional states, and learning responses. Nevertheless, this task is
difficult due to the intricate and highly variable nature of face expressions. Traditional approaches often fall short
in accurately detecting and interpreting subtle facial cues, resulting in limitations in effectively monitoring
student learning. The growing interest in affective computing has led to the development of systems capable of
recognizing and responding to emotional cues [5]. Among such applications is Face Expression Recognition
(FER), which has become increasingly relevant in educational contexts. FER provides feedback that can enhance
instructional adaptation, improve learner satisfaction, and support personalized learning [6], [7]. In recent years,
FER also has gained traction in applications such as education, security systems, healthcare diagnostics, and
customer experience analysis [7], [8].
Developing a robust FER model often involves complex deep learning architectures and large-scale annotated
datasets. One of the solutions is by applying Google Teachable Machine (GTM) to create the FER model. These
challenges can be solved comprehensively specifically for educator. GTM simplify this process by offering an
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 866
www.rsisinternational.org
intuitive, no-code interface that leverages pre-trained models for transfer learning [9]. GTM provides a simple
and efficient approach for developing a stable FER model that can continuously monitor students' emotional
states and engagement levels in real time. This approach helps close the gap in face expression analysis by
providing educators with meaningful insights into student learning behaviors, enabling timely interventions and
personalized support to improve the online learning experience. GTM offers an accessible alternative, enabling
non-technical users to create classification models with ease. This makes it particularly attractive for educators
and researchers looking for rapid deployment in real-world learning environments. GTM accepts three types of
input from users which are image, audio and pose. The inputs can be gathered via webcam, microphone or upload
[9].
Therefore, this study aims to develop a FER model using GTM to support learning feedback in higher education.
The desired output is building an accurate, fast model that can precisely categorize face expression into five
different classes which are happiness, sadness, surprise, anger and neutral. The dataset will be obtained from an
online repository. The developed model has potential applications in virtual classrooms, online assessments, and
self-paced e-learning environments.
The remainder of this paper is organized as follows. Section II reviews related work on FER while section III
describes the methodology of the research starting from data preparation until deployment. Section IV presents
the experimental result and discussion. Finally, section 6 concludes the research together with limitations and
directions for future research.
LITERATURE REVIEW
FER systems traditionally rely on feature extraction and machine learning classification. Early approaches used
hand-crafted features such as Scale-Invariant Feature Transform (SIFT) and Local Binary Pattern (LBP) [10].
The rise of deep learning led to widespread use of convolutional neural network (CNN) [11]. Ko (2018) provided
a review of FER technologies and noted that machine learning techniques such as CNN and Long Short-Term
Memory (LSTM) are highly effective in emotion classification tasks. There are many researches related to FER
have been done. For example, Whitehill et al. [12] developed an automated FER system to monitor student
engagement. Their system used a CNN to predict attention levels and found a strong correlation with academic
performance. Minaee et al., [13] conducted experiments on four different datasets using their proposed end-to-
end deep learning framework based on an attentional convolutional network. From their experiment, by using
the FERG dataset, they attained the highest accuracy rate of around 99.3% compared to using other datasets.
More interestingly, many researchers using hybrid techniques or combined several techniques for FER. In
example, Abinaya et al. [14] proposed Hybrid Adaptive Kernel based Extreme Learning Machine (HAKELM)
scheme on their research and achieved 95.5% of accuracy, 90.12% of sensitivity, and 95.1% specificity compared
to the previous existing algorithm. Rahul et al. [15] proposed a hybrid approach for emotion recognition by
combining CNN and Recurrent neural networks (RNN) and tested using three datasets. FER-13 dataset achieved
94.08% of accuracy rate meanwhile EMOTIC dataset attained 72.64% and the lowest accuracy rate is 68.10%
for FERG dataset. Moreover, Kong et al. [16] introduced a real-time FER method utilizing iterative transfer
learning and an Efficient Attention Network (EAN), specifically designed for edge environments with limited
resources. This approach effectively addresses issues related to server overload and the risk of privacy breaches.
Developing a robust FER model often involves complex deep learning architectures and large-scale annotated
datasets. Recent studies have explored using MobileNet, Visual Geometry Group (VGG), and Residual Network
(ResNet) for lightweight deployment in resource-constrained environments. MobileNet has shown good trade-
offs between performance and efficiency [17]. For example, Haslini et al. [18] developed android application for
FER using Personal Image Classifier, where their backend engine using MobileNet. Besides, Aly [19] utilized
ResNet50+CBAM+TCNs to track student engagement in online classrooms. The proposed techniques achieved
accuracies of 91.86% for RAF-DB, 91.71% for FER2013, 95.85% for CK +, and 97.08% for KDEF dataset.
Moreover, Huang et al. [20] used six emotion categories related to classroom teaching and learning and their
results showed that MultiEmoNet achieved a classification accuracy of 91.4% on a homemade classroom student
emotion dataset. Gao et al. [8] highlighted classroom expression recognition systems using spatial, channel, self-
attention for teaching feedback. They constructed five category of classroom expressions and the proposed
method got 88.34% of accuracy in expression recognition tasks and offers strong support for smart classrooms.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 867
www.rsisinternational.org
Table I summarized several current researches on FER that consist information on number of face expression,
dataset and techniques. Based on the table, many researchers used custom FER dataset and also several
benchmark datasets for FER research including FER2013, Radboud Faces Database (RAF-DB), and Extended
Cohn‑Kanade (CK+). These datasets have enabled researchers to develop models with high accuracy [13], [16].
FER2013 is a large-scale dataset collected from the internet, consisting of face images extracted from YouTube
videos. It comprises of seven emotion categories: anger, disgust, fear, happiness, sadness, surprise, and neutral
[19]. RAF-DB is a dataset that consists of static images of face expressions collected from online sources. It
comprises a variety of emotion categories such as happiness, surprise, fear, disgust, sadness, and anger. CK +
dataset consists of posed face expressions of emotions such as happiness, sadness, anger, surprise, disgust, and
fear. All three datasets are popular in FER research. In term of techniques, most of the researchers using deep
learning technique to classify face expression such as CNN, RNN, ResNet and YOLO. In addition, although
Ekman’s emotion theory is a landmark in the field of emotion recognition, its categorization may not fully
capture the complexity of students’ emotions in a specific classroom setting [20]. However, due to publicly
available databases of students’ emotions are limited [21] and student privacy issues, many researchers still used
seven basic categories of face expression which are anger, disgust, fear, happiness, sadness, surprise, and neutral
[19], [22].
The importance of accessible artificial intelligence (AI) tools like GTM is rising. GTM provides an intuitive way
to build image classifiers without coding and widely used in education to teach machine learning (ML) concepts
[23]. Carney et al., [23] emphasized several benefits of GTM, including its user-friendly interface, the absence
of any requirement for coding or prior ML experience, and its potential to offer interactive tools and simplified
concepts that make teaching and learning ML more accessible. In other words, it enables individuals from various
backgrounds to use ML without requiring specialized knowledge or technical skills. In addition, Wong & and
Fadzly [24], highlighted that while GTM does not offer deep customization, it is highly effective for rapid proof-
of-concept models and training with small datasets. Such tools are becoming increasingly relevant as interest in
low-code and no-code AI development grows.
Table I: Several researches on face expression recognition
Researcher, Year
Number of Expression
Dataset
Technique
[8], 2025
5
Custom FER dataset
Multi-attention fusion network (MAF-ER)
[19], 2024
7
RAF-DB, FER2013,
CK + and KDEF
ResNet50+CBAM+TCNs
[22], 2024
7
FER2013, FERPlus,
RAF-DB, AffectNet,
real smart classroom
facial expression
dataset (SCFED)
Multi-scale and deep fine-grained feature
attention enhancement (MDFAE)
[20], 2024
6
Custom FER dataset
Enhance YOLOv8
[18], 2022
3
Custom FER dataset
Personal Image Classifier (CNN)
[16], 2022
7
FER2013, RAF-DB
EAN
[15], 2022
7
EMOTIC, FER-13,
FERG
CNN and RNN
[14], 2021
7
AT&T, YALE FACE B
HAKELM
[13], 2021
7
FER2013, CK+,
JAFFE, FERG
End-to-end deep learning framework based
on an attentional convolutional network
Therefore, this paper proposes development of FER model using GTM to perform multiclass student face
expression classification in higher education context.
METHODOLOGY
This section consists of information related to the methodology for developing FER model using GTM which
are hardware requirement, data preparation, model configuration, training process and finally model testing and
validation.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 868
www.rsisinternational.org
Hardware and Software Requirement
To develop the FER model using GTM, a personal computer (PC) was utilized with the following specifications:
an AMD Ryzen 5 4500 6-core processor, 8 GB of RAM, and Windows 10 Pro as the operating system. Google
Chrome, along with a stable internet connection was used to access the GTM platform and perform the model
training process.
Data Preparation
The model training process begins with data collection, where images representing different face expressions are
collected from RAF-DB dataset [25] and five classes has been created: happiness, sadness, surprise, anger and
neutral. Each class is represented by multiple images to improve the model's ability to generalize and recognize
expressions across numerous conditions. Preprocessing involves in this step which is cleaning activity to remove
blur images and also children’s images. A total of 600 images were selected and then uploaded into GTM
website. Fig. 1 represents some sample images from RAF-DB dataset.
Fig. 1. Sample images of dataset
For this study, we used a single holdout method in GTM to split data for training and testing/validation. This
split is standard machine learning practice to prevent overfitting. 85% of the total data is used for training (510
images), while 15% is reserved for internal testing and validation (90 images) as shown in Table II. This ratio is
commonly used to balance model training and validation [26].
TABLE II: DATA SPLITTING USING HOLDOUT METHOD
Class
Training (85%)
Testing / Validation(15%)
Total
Happiness
102
18
120
Sadness
102
18
120
Surprise
102
18
120
Anger
102
18
120
Neutral
102
18
120
Total
510
90
600
Model Configuration
In supervised learning, particularly deep learning, several hyperparameters control the learning process which
are epoch, learning rate and batch size. Table III shows the parameter setting for training the FER model using
GTM. These parameter values are usually used in small image classification tasks using MobileNet [17], [27].
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 869
www.rsisinternational.org
Table III: Training parameter
Parameter
Value
Epoch
50
Learning rate
0.001
Batch size
16
Epochs are defined as one complete forward and backward pass of all training samples. Each epoch allows the
model to learn and refine its parameters through weight updates. Too few epochs may result in underfitting,
while too many may lead to overfitting [26]. In this study, suitable value for epoch is 50.
Learning rate controls the size of the steps the model takes to update its weights during the training process. A
high learning rate may lead to fast convergence but overshooting, while a low rate results in slower, more stable
learning [28]. In this study, a low learning rate, 0.001 is used to make sure the stable learning process for creating
FER model.
Batch size refers to the number of data samples processed before the model’s parameters are updated. Smaller
batches provide more updates but may be noisy, while larger batches are computationally efficient but less
flexible [29]. For this study, a smaller value of batch size which is 16 is applied for training purpose.
Training Process
GTM uses Tensorflow.js, an open source machine learning library in JavaScript to train and run the training
result in a model in a web browser [30]. GTM also leverages the concept of transfer learning where instead of
training a neural network from scratch, it uses a pre-trained MobileNet model. Transfer learning has proven
effective, allowing pre-trained models to be fine-tuned on emotion classification tasks. MobileNet is a CNN with
a smaller model size with less trainable parameter and calculation amount [18].
Therefore, the training process in GTM consists of the following steps:
1. The uploaded face expression images were pre-processed which included image resizing and
normalization. Images were automatically resized to 224x224 pixels by GTM.
2. A CNN backbone is used internally to learn distinguishing features of each face expression.
3. The model is trained using transfer learning, where a pre-trained model (MobileNet) is fine-tuned on the
new dataset.
4. The model continues training for a set number of epochs until it achieves satisfactory accuracy, evaluated
via loss and accuracy metrics.
Model Testing and Validation
Once the training is completed, the model is validated using the test dataset (15%). This helps assess the model’s
generalization ability. A confusion matrix as shown in Fig. 2 is a table that summarizes the classification results,
showing the number of true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN)
for each class. Confusion matrix displays the number of predictions made by the algorithm compared to the
actual true values in the test dataset. The Y axis (Class) denotes to the class of sample, while the X axis denotes
to the predicted class.
Fig. 2 Multiclass classification in confusion matrix [31]
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 870
www.rsisinternational.org
Moreover, based on the result from confusion matrix, four key metrics that commonly utilized to assess the
model’s effectiveness are Accuracy, Precision, Recall, and F1-Score [32], [33]. Accuracy refers to overall
correctness of the model or the proportion of correctly classified face expressions to the total number of
expressions [19]. Precision measures how many predicted positives are actually correct. Recall, also known as
Sensitivity measures how many actual positives were correctly predicted. F1-Score is the harmonic mean of
Precision and Recall, providing a balanced assessment of the system's performance. Therefore, the formula for
Accuracy, Precision, Recall, and F1-Score are displayed in Eq.1, Eq.2, Eq.3 and Eq.4 respectively. TP denotes
True Positives (correctly recognized expressions), TN signifies True Negatives (correctly ignored expressions),
FP represents False Positives (incorrectly recognized expressions), and FN indicates False Negatives (incorrectly
ignored expressions) [19].
(1)
(2)
(3)
(4)
RESULT AND DISCUSSION
Table IV shows the accuracy result per class. Based on the table, the highest accuracy result is Neutral (0.86 or
86%) followed by Surprise (0.89 or 89%). Both Happiness and Anger have the same accuracy of 0.83 or 83%.
The lowest accuracy class is Sadness (0.72 or 72%). This result shows that Neutral face expression is the easiest
to be recognized compared to Sadness face expression that is the most difficult to be identified.
TABLE IV: ACCURACY PER CLASS
Class
Accuracy (%)
#Samples
Happiness
83
18
Sadness
72
18
Surprise
89
18
Anger
83
18
Neutral
94
18
Fig. 3 shows the result for confusion matrix, accuracy per epoch and loss per epoch from GTM internal analysis.
Based on the confusion matrix in Fig. 3(a), in general result, the FER model has performed strongly with most
predictions along the diagonal are correct classification. Misclassifications occur in the FER model but are
relatively small in number. The best recognized classes are Neutral and Surprise while the most confused class
in Sadness. In example, class Neutral is very high accuracy because this expression is consistently recognized
with one misclassified as Sadness.
(a) Confusion Matrix
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 871
www.rsisinternational.org
(b) Accuracy per epoch
(c) Loss per epoch
Fig. 3 Result for confusion matrix, accuracy per epoch and loss per epoch
Based on confusion matrix in Fig. 3(a), the value of accuracy, precision, recall and F1-Score can be calculated.
Table V displays the summary result for precision, recall and F1-Score for each class. Based on the calculation,
the FER model achieved an overall accuracy of 84.44%, demonstrating good performance for five classes. The
model showed its strongest results for Neutral and Happiness, both achieving high precision and F1-scores. In
contrast, Anger and Sadness were more challenging for the model, with lower precision and recall due to
confusion between the two classes. This is expected in FER studies, as negative emotions often share similar
facial cues. Despite this, the model still reached acceptable performance levels for these classes.
TABLE V: RESULT FOR PRECISION, RECALL AND F1-SCORE
Class
Precision (%)
Recall (%)
F1-score (%)
Happiness
100
83.33
90.91
Sadness
76.47
72.22
74.29
Surprise
84.21
88.89
86.49
Anger
71.43
83.33
76.92
Neutral
94.44
94.44
94.44
The result for accuracy per epoch is depicted in Fig. 3(b). Accuracy reflects how correctly the prediction model
performs. It represents the percentage of correct classifications made during training. A perfect prediction yields
an accuracy of one, while any errors result in a value less than one. A good accuracy is shown by the intercept
line between actual accuracy and the test accuracy. Moreover, the result for lost per epoch is shown in Fig. 3(c).
Loss per epoch represents the number of errors during each training cycle (epoch); generally, a lower loss value
indicates better model performance.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 872
www.rsisinternational.org
Moreover, the performance comparison between our proposed FER model and several other researchers using
the RAF-DB dataset [16], [19], [22] is shown in Table VI. The proposed FER model’s accuracy is comparable
to more complex deep learning-based models, but achieved using a simpler tool with no coding involved and
smaller size dataset.
TABLE VI: PERFORMANCE COMPARISON OF OTHER RESEARCHERS ON RAF-DB DATASET
Researcher
Method
Accuracy (%)
[22]
Multi-Scale and Deep Fine-Grained Feature Attention
92.93
[16]
EAN
85.30
[19]
ResNet50, CBAM, and TCNs
91.86
Proposed
GTM
84.44
Furthermore, each face expression has the specific learning feedback interpretation as summarized in Table VII.
For example, expression Happiness means positive learning engagement, learner understands the learning
content or learner feels satisfaction and motivated. Sadness expression interprets that learner feels boring with
the content, learner needs emotional support or fatigue during the classroom session.
TABLE VII: RELATIONSHIP BETWEEN FACE EXPRESSION AND LEARNING FEEDBACK
Class
Learning feedback interpretation
Happiness
Engagement, understanding, satisfaction
Sadness
Boredom, demotivation, fatigue
Surprise
Attention, curiosity, cognitive shift
Anger
Frustration, cognitive overload
Neutral
Focused attention, passive engagement
This FER model can be integrated into online learning platforms or physical classrooms via webcams to provide
real-time emotion monitoring in higher education environment. For instance, if a learner shows Sadness emotions
over time, the system could prompt a motivational message or offer help materials. Educators could use the
insights to adjust content delivery dynamically.
CONCLUSION
This research demonstrates the successful development of a multiclass face expression classification model using
GTM. The proposed model reliably distinguishes between five expressions. With an overall accuracy of 84.44%,
this model is effective in distinguishing key facial expressions for multi-class setup. The proposed FER model
can be exported in multiple formats such as TensorFlow.js (for web-based deployment), TensorFlow Lite (for
mobile and embedded systems) and downloadable Keras model (for further development). The FER model can
be integrated into applications or systems that provide real-time learning feedback, classroom monitoring, or
affective computing systems, making it valuable in educational technology research.
One of the primary contributions of this research is the use of GTM to simplify FER model training. This
approach enables educators to create custom feedback models without any programming knowledge, low setup
requirement, and intuitive interface making it a suitable solution for real-world educational applications.
Some limitations of the model include limited generalization across diverse datasets and model overfitting due
to small dataset. Future work will focus on expanding the model using larger datasets with real student face
expressions and embedding the model in Learning Management Systems (LMS).
REFERENCES
1. R. Pekrun, “The Control-Value Theory of Achievement Emotions: Assumptions, Corollaries, and
Implications for Educational Research and Practice,” Educ. Psychol. Rev., vol. 18, no. 4, pp. 315–341,
2006.
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 873
www.rsisinternational.org
2. P. Ekman and W. V Friesen, “Constants across cultures in the face and emotion.,” Journal of Personality
and Social Psychology, vol. 17, no. 2. American Psychological Association, US, pp. 124–129, 1971.
3. Y. Huang, F. Chen, S. Lv, and X. Wang, “Facial Expression Recognition: A Survey,” Symmetry (Basel).,
vol. 11, no. 10, 2019.
4. A. Mehrabian, “Communication without words,” in Psychology Today, 1968, pp. 51–52.
5. R. W. Picard, Affective Computing. MIT Press, 2000.
6. R. A. Calvo and S. D’Mello, “Affect Detection: An Interdisciplinary Review of Models, Methods, and
Their Applications,” IEEE Trans. Affect. Comput., vol. 1, no. 1, pp. 18–37, 2010.
7. S. D’mello and A. Graesser, “AutoTutor and affective autotutor: Learning by talking with cognitively
and emotionally intelligent computers that talk back,” ACM Trans. Interact. Intell. Syst., vol. 2, no. 4,
Jan. 2013.
8. Y. Gao, L. Zhou, and J. He, “Classroom Expression Recognition Based on Deep Learning,” Appl. Sci.,
vol. 15, no. 1, 2025.
9. “Google Teachable Machine,” 2025. [Online]. Available: https://teachablemachine.withgoogle.com/.
[Accessed: 20-May-2025].
10. C. Shan, S. Gong, and P. W. McOwan, “Facial expression recognition based on Local Binary Patterns:
A comprehensive study,” Image Vis. Comput., vol. 27, no. 6, pp. 803–816, 2009.
11. A. Mollahosseini, D. Chan, and M. H. Mahoor, “Going deeper in facial expression recognition using
deep neural networks,” in 2016 IEEE Winter Conference on Applications of Computer Vision (WACV),
2016, pp. 1–10.
12. J. Whitehill, Z. Serpell, Y. C. Lin, A. Foster, and J. R. Movellan, “The faces of engagement: Automatic
recognition of student engagement from facial expressions,” IEEE Trans. Affect. Comput., vol. 5, no. 1,
pp. 86–98, 2014.
13. S. Minaee, M. Minaei, and A. Abdolrashidi, “Deep-Emotion: Facial Expression Recognition Using
Attentional Convolutional Network,” Sensors, vol. 21, no. 9, 2021.
14. D. Abinaya, C. Priyanka, M. Rocky Stefinjain, G. K. D. Prasanna Venkatesan, and S. Kamalraj,
“Classification of Facial Expression Recognition using Machine Learning Algorithms,” J. Phys. Conf.
Ser., vol. 1937, no. 1, p. 12001, Jun. 2021.
15. M. Rahul, N. Tiwari, R. Shukla, D. Tyagi, and V. Yadav, “A New Hybrid Approach for Efficient Emotion
Recognition using Deep Learning,” Int. J. Electr. Electron. Res., vol. 10, no. 1, pp. 18–22, 2022.
16. Y. Kong, S. Zhang, K. Zhang, Q. Ni, and J. Han, “Real-time facial expression recognition based on
iterative transfer learning and efficient attention network,” IET Image Process., vol. 16, no. 6, pp. 1694–
1708, 2022.
17. A. G. Howard et al., “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision
Applications,” CoRR, vol. abs/1704.0, 2017.
18. S. Haslini, A. Hamid, N. Mustapa, and M. F. Mustapha, “An Android Application for Facial Expression
Recognition Using Deep Learning,” J. Appl. Math. Comput. Intell., vol. 11, no. 2, pp. 505–520, 2022.
19. M. Aly, “Revolutionizing online education: Advanced facial expression recognition for real-time student
progress tracking via deep learning model,” Multimed. Tools Appl., vol. 84, no. 13, pp. 12575–12614,
2024.
20. Y. Huang, W. Deng, and T. Xu, “A Study of Potential Applications of Student Emotion Recognition in
Primary and Secondary Classrooms,” Appl. Sci., vol. 14, no. 23, 2024.
21. B. Fang, X. Li, G. Han, and J. He, “Facial Expression Recognition in Educational Research From the
Perspective of Machine Learning: A Systematic Review,” IEEE Access, vol. 11, pp. 112060–112074,
2023.
22. Z. Shou et al., “A Student Facial Expression Recognition Model Based on Multi-Scale and Deep Fine-
Grained Feature Attention Enhancement,” Sensors, vol. 24, no. 20, 2024.
23. M. Carney et al., “Teachable Machine: Approachable Web-Based Tool for Exploring Machine Learning
Classification,” in Extended Abstracts of the 2020 CHI Conference on Human Factors in Computing
Systems, 2020, pp. 1–8.
24. J. J. N. Wong and N. and Fadzly, “Development of species recognition models using Google teachable
machine on shorebirds and waterbirds,” J. Taibah Univ. Sci., vol. 16, no. 1, pp. 1096–1111, Dec. 2022.
25. Dev-ShuvoAlok, “RAF-DB DATASET,” 2023. [Online]. Available:
https://www.kaggle.com/datasets/shuvoalok/raf-db-dataset?select=DATASET. [Accessed: 01-May-
2025].
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025
Page 874
www.rsisinternational.org
26. I. Goodfellow, Y. Bengio, and C. Aaron, Deep Learning. MIT Press, 2016.
27. M. Tan and Q. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” in
Proceedings of the 36th International Conference on Machine Learning, 2019, vol. 97, pp. 6105–6114.
28. L. Bottou, “Stochastic Gradient Descent Tricks,” in Neural Networks: Tricks of the Trade: Second
Edition, G. Montavon, G. B. Orr, and K.-R. Müller, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg,
2012, pp. 421–436.
29. D. Masters and C. Luschi, “Revisiting Small Batch Training for Deep Neural Networks,” CoRR, vol.
abs/1804.0, 2018.
30. E. Malahina, R. Hadjon, and F. Bisilisin, “Teachable Machine: Real-Time Attendance of Students Based
on Open Source System,” IJICS (International J. Informatics Comput. Sci., vol. 6, no. 3, p. 140−146,
2022.
31. K.-A. Tait et al., “Intrusion Detection using Machine Learning Techniques: An Experimental
Comparison,” in 2021 International Congress of Advanced Technology and Engineering (ICOTEN),
2021, pp. 1–10.
32. A. S. A. Mohammed Aly, “EMU-Net: Automatic Brain Tumor Segmentation and Classification Using
Efficient Modified U-Net,” Comput. Mater. \& Contin., vol. 77, no. 1, pp. 557–582, 2023.
33. M. H. Behiry and M. Aly, “Cyberattack detection in wireless sensor networks using a hybrid feature
reduction technique with AI and machine learning methods,” J. Big Data, vol. 11, no. 1, p. 16, 2024.
