With the pervasive application of the new generations of
mobile communications, computer networking, Internet of
Things (IoT), cloud computing, and artificial intelligence
(AI), the world is delving into a new era of intelligent
information technology and large scale of data mining
applications. Artificial Intelligence, machine learning, and
big data processing are increasingly becoming the main
motives for digital transformation all over the world and
control all aspects of our lives [1], [2]. These revolutionary
developments in information technology resulted in new
methodologies in teaching and learning processes at higher
education institutions. Artificial Intelligence helps in the
enrichment of students’ theoretical background and
enhances their university education in terms of diversifying
the development of their educational needs, taking into
consideration the huge impact of using the Internet, and the
wide opportunities to experiment new things and more
complex behavior patterns. Moreover, AI, networking, and
big data help in providing positive educational and cultural
recommendations to users in real time, which can assist in
meeting the expanded development requirements of
university students [3]. On the other hand, employment of AI
in higher education provides universities and colleges with
the ability to create a friendly relationship with students, and
helps in improving, to some extent, the identity and self-
confidence of university students, besides developing
harmonious relationships throughout the campus
environment in the new educational system.
The optimization of the relationship between students
and teachers and the promotion of friendly collaboration
between different departments at the university lead to the
creation of healthy atmosphere throughout the whole campus
and establishing the basic requirements of ethical and moral
education at the university. The establishment of morality
lies in encouraging students to apply right moral principles
and qualities and to apply distinguished moral reality.
Therefore, it is possible for the colleges and universities to
help students to be more competent and develop desirable
qualities by applying various educational methods, and to be
main facilitators, participants, and promoters for the students
under the enormous wave of big data which includes wide
range of data resources, advanced data processing, and usage
of artificial intelligence in this regard. Due to its countless
advantages and benefits, artificial intelligence has created
new approaches to big data acquisition and data processing,
besides enabling the usage of new educational methods to
promote innovation of thinking, reconstruction of
paradigms, and implementing scientific and logical solutions
for daily life problems. The role of artificial intelligence and
deep learning in improving higher education resulted in
establishing robust tools to fix the weaknesses and
shortcomings in educational systems such as cold starts, data
sparsity, and deprived interpretability.
The introduction of Convolutional Neural Networks
(CNN) and Recurrent Neural Networks (RNN) enabled
linguistic processing tasks, such that students can practice
various deep learning approaches such as Deep Cooperative
Neural Networks (DeepCoNN) to excavate user preferences
and suggest recommendations from the user’s perspective
[4], [5]. DeepCoNN is primarily based on CNN model, and
it consists of two parallel neural networks which learn the
students’ behavior and the learning outcomes of the
university course, respectively, and creates the necessary
connections between them at the top of the network to
emphasize the interaction and attention mechanisms, which
is used for suggesting recommendation algorithms that
improve the learning/teaching process at higher education
institutions. Neural Attentional Rating Regression (NARRE)
employs attention mechanisms to comprehend textual
depictions to better understand users’ behaviors, predict
users’ interests, and create proper interpretations, based on
two-channel attention mechanism. On the other hand, the
advent of Neural Language Processing (NLP) helped in the
development of text in the area recommendation.
Unfortunately, NLP is unidirectional which limits the power
of interpretation. Hence, a bidirectional pertaining model
with a high generalization capability called BERT was
proposed in [26], which reads the whole text at once using
A Proposed Artificial Intelligence Algorithm for Development of
Higher Education
AMIN AL KA’BI
Australian University, KUWAIT
Abstract: Higher education has delved into a new stage of rapid development focusing on quality improvement, while
encountering new challenges and obstacles. In this research work, an artificial intelligence algorithm for education
improvement is proposed. Firstly, deep feature abstraction in temporal and special dimensions is performed using Long
Short-Term Memory (LSTM) artificial neural network and convolutional networks. Consequently, multiscale attention
fusion techniques are used to improve the articulateness of the characteristics and come up with better recommendations
with the assistance of multilayer perceptron. Moreover, the proposed model helps in improving the cognitive capability of
students and enhances their overall quality of perception. Moreover, it has been proven that the performance of the proposed
model provides better recommendation outcomes and better robustness compared to existing models through conducting
extensive experiments based on real data.
Keywords: Artificial Intelligiance (AI); Communication Sysems; Higher Education; Convolutional Neural Networks
(CNN); Recurrent Neural Networks (RNN).
Received: February 19, 2022. Revised: November 15, 2022. Accepted: December 9, 2022. Published: January 19, 2023.
1. Introduction
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
7
Volume 22, 2023
Transformer’s Encoder, and this procedure allows the
algorithm to learn on both sides of the word and hence digest
the meaning of the word in the sentence more accurately and
provides a good ground for downstream tasks [6-9].
This paper investigates various techniques used in deep
learning such as CNN and LSTM to suggest proper
recommendation models to be employed in higher
educational institutions, which helps the students in locating
their knowledge of interest and real attention. Moreover, it
helps in improving the cognitive capability of the students
and the overall quality of their mindsets.
The proposed model of higher education presented in this
paper is supported by artificial intelligence and it is based on
convolutional neural networks (CNN) and progressive
features obtained from the progressive dimension by using
bidirectional neural networks. Sec. 2.1 explains the theory of
CNN, and the procedure of feature extraction from
progressive dimension is introduced in Sec. 2.2. The scale
fusion attention is explained in Sec. 2.3. The procedure of
calculating the prediction scores and suggesting
recommendations is explained in Sec. 2.4.
The Convolutional Neural Networks (CNN) Model
consists of an input layer, pooling layer, fully connected
layer, convolutional layer, and an output layer [10], [11]. In
this model, the convolution layer is the outcome of the inner
product of the filter matrix and the embedding vector. The
pooling layer groups each feature map acquired from the
convolution layer. It is found that the better features can be
extracted using the maximum pooling approach in
comparison with the mean pooling approach. The output
features of the pooling layer are taken by the fully connected
layer as an input and activated by the activation function to
generate a fixed dimension feature vector. The performance
of CNN is excellent in dealing with classification problems;
however, its recommendation algorithms produce small
number of results. This is due to the fact that the
recommendation process is a regression process, and they
have different objectives [12], [13]. To obtain more precise
recommendation results for text data, it is possible to build a
hybrid recommendation model by merging a classical
recommendation algorithm with CNN. Fig. 1 depicts the
hierarchy of the Convolution neural network (CNN) model.
2.1.1. Input Layer.
In CNN model structure, the role of the input layer is to
transform the textual data of the learning resource into an
embedding matrix, where the dimensionality of the
embedded data is very low, with the possibility of converting
the discrete data sequences into continuous data vectors.
Fig. 1. Convolution neural network (CNN) model.
Equation 1 illustrates the embedding matrix D where
each row in the matrix represents a clause element.
where m denotes the embedding dimensionality, n denotes
the number of the words, and [wi,1: m] denotes the vector
representation of the i-th word.
2.1.2. Convolution Layer.
The convolution operations are performed on the
embedding matrix by using multiple convolution kernels
with different dimensions. To cover the entire word
embedding vector discussed in this paper, the size of
convolution is determined by product of the vector
dimension and the number of words. Fig. 2 illustrates the
flow of the convolution operation.
Fig. 2. Flow of the convolution operation.
2.1.3. Pooling Layer.
The main purpose of this layer is to decrease the
dimension of the feature map and to decrease the number of
parameters used in the network. The main pooling operations
2. Methodology
2.1. Convolutional Neural Networks
(CNN) Model
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
8
Volume 22, 2023
include maximum pooling and average pooling. In the
pooling operation, little deviations in the feature maps can be
discarded and hence the accuracy can be improved while
averting the phenomenon of overfitting. If the feature map
acquired in the t-th convolutional layer is
, then the maximum value of can be
obtained by using the maximum pooling strategy as
illustrated in Eq. (2).
(2)
where expresses the pooling outcome of the ti-th
convolutional layer.
2.1.4. Fully Connected Layer. The function of the fully
connected layer is to implement the feature vector and the
extracted layer values. Assume that the fully connected layer
has m neurons, and a text feature vector is created after
initiating the ReLu activation function such as:
(3)
where denotes the ReLu activation function, is the
output of the learning resource text information on the
pooling layer, is the bias, and denotes the weight.
The neurons of the Long Short-Term Memory (LSTM)
artificial neural network accept only the data from the
neurons close to the layers, meanwhile the words before and
after affect the semantic connections. However,
Bidirectional Long Short-Term Memory Recurrent Neural
Network (BiLSTM-RNN) consists of two groups of long and
short term recurrent neural networks with contradictory
learning trends, which enables better understanding of the
related semantics in comparison with LSTM. The LSTM is
composed mainly from four elements input gate ,
forgetting gate
memory unit , and output gate The
input gate controls the flow of the data into the memory unit,
the forgetting gate identifies the holding of the previous state
data in the memory unit, which determines the memory state
based on the current input data; and then the output gate
identifies the output value of the memory unit for the next
state. The relevant computation procedure is illustrated in
Eq. 4 to Eq. 9.
(4)
(5)
(6)
(7)
(8)
(9)
where represents the memory cell state, is the data
input, b is the function bias term, W denotes the matrix
multiplication operation, is the dot product operation, and
is the sigmoid function.
BiLSTM-RNN model consists of two groups of forward and
backward LSTM models attached to a learning fearure
denoted by
and
respectively. Eq. 10
expresses the time-dimensional feature which determines
final representation of the LSTM model structure.
(10)
where denotes the concatenation operator. This operation
enables the BiLSTM model to fully process the input words
of the contextual data.
The spatial dimensional features F and the time-dimensional
feature T can be merged as depicted in Fig. 3, where
multiscale feature synthesis attention process is used to
achieve this goal.
Fig. 3. Multiscale feature synthesis attention process.
Firstly, the matrix FA and the matrix FB which
represent the matching between the dimensional features
denoted by Fi and the attribute features denoted by Ti are
determined as expressed in Eq. 11.
& (11)
Secondly, the function SoftMax is used to find the
attention distribution weights w1 and w2 of the matching
matrices. Then, the attention representation matrices F′i
and T′I are calculated by multiplying the weights w1 and w2
with the individual scale features as expressed in Eq. 12.
& (12)
where (×) denotes matrix fork multiplication.
At last, the inter-scale mutual attention matrices F1 and
F2 are calculated by using a multiplicative gating process
to multiply the attentional representation with another
single-scale feature for the relevant elements as expressed
in Eq. 13.
2.2. Long- and Short-time Bidirectional
Recurrent Neural Network
2.3 Multiscale Feature Fusion
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
9
Volume 22, 2023
& (13)
where (·) denotes matrix dot product
To determine the final multi-scale synthesis features, the
functions F1 and F2 are operated as expressed in Eq. 14.
(14)
where is the Cat operation.
The outcome of Eq. 14 is used as an input to the
multilayer perceptron the anticipate the scores. A cross-
entropy loss function is used to obtain the end-to-end
optimization of the model, and to find and to anticipate the
weights for each layer. Eventually, anticipation outcomes are
determined by relating the activation function to the range
[0, 1]. Equations 15-17 show the corresponding calculations.
(15)
where Xt is the fully connected outcome and ht is the
output hidden vector of the decoder.
(16)
where y is the true description, x is the fully connected, and
P is the SoftMax function.
(17)
where θ denotes the parameter of cross-entropy balance
loss of the model.
Experiments were carried out using PyTorch deep
learning framework, NVIDIA CUDA 11.3, and cuDNN
deep learning acceleration library. Stochastic gradient
descent algorithm SGD was used to train the network with a
momentum factor of 0.8, learning decay rate of 0.001, and
an initialized learning rate of 0.005. Fig. 4 depicts the
accuracy curve and the model training loss curve.
A random removal of some neurons is necessary to
overcome the overfitting problem. To achieve this goal a
Dropout of 0.5 is used. Fig. 4 shows that when the iteration
number approaches 40, both loss curves of test and training
get smoother, and the amount of loss decreases below 0.06
which is an indication of the model convergence.
3.1.1. Proof of model validity. Experiments are carried out to
prove the model validity using identical environmental and
evaluation indices. the Gram-based modifieded
recommendation algorithm Gram-CF, the traditional
collaborative filtering recommendation algorithm User-CF
based on user play records, and the collaborative filtering
recommendation algorithm FCNN-CF based on user
preference statistics are chosen as reference models. The
evaluation metrics used are the F1 value, precision rate (PR),
accuracy rate (AR), and recall rate (RR).
Fig. 4. Loss and accuracy in testing and training and phases.
(a) Loss, (b) Accuracy.
Fig. 5 exhibits a significant improvement of the proposed
model in this paper by 1.78%, 1.22%, and 3.05% in terms of
accuracy in comparison with the reference models FCNN-
CF, User-CF, and Gram-CF recommendation models,
respectively.
Fig. 5. Comparison of the recommended performance of different
models.
2.4. Scores Prediction and
Recommendations Suggestion
3. Experimental Results
3.1. Experimental Results
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
10
Volume 22, 2023
However, the proposed model shows and improvement
in terms of precision by 2.60%, 0.78% and 1.79%,
respectively compared to the reference models, and it shows
an improvement in terms of recall, by 1.67%, 1.10%, 91.6%,
and 2.58%, respectively. Meanwhile, the improvement in
terms of F1 is 1.58, 2.38%, and 0.22%, respectively.
Consequently, these experimental results prove that the
proposed model shows better performance at all aspects
compared to the reference models: FCNN-CF, User-CF, and
Gram-CF. This improvement is due to two-dimensional
feature extraction, so that temporal and spatial are merged to
reinforce the features expressiveness, which in turn removes
anomalous data, and improves the learning capability of the
model.
3.1.2. Robustness Verification. The robustness of the
proposed model is verified by performing experiments under
similar input data and similar experimental platform. The
results of the experiments are depicted in Fig. 6. It can be
concluded that the robustness of the proposed model is
higher than 81% for AR, PR, RR, and F1 metrics compared
to the reference models as the number of recommendations
approaches 25, which verifies the robustness of the proposed
model.
Fig. 6. Recommendation performance versus the number of
recommendation items.
3.1.3. Real-Time Verification of Recommendation algorithm.
The test results of the real-time performance of the
recommendation algorithm of the proposed model are
illustrated in Fig. 7, where tests are performed in the same
environments with the same data.
Fig. 7: Success rates of various recommendation reference models.
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
11
Volume 22, 2023
Fig. 7. Shows that the proposed model has the highest
recommendation rate where it has a recommendation rate of
6.9 video/s, FCNN-CF model has a recommendation rate of
6.1 video/s, the User-CF model has a recommendation rate
of 6.4 video/s, and the Gram CF model has a
recommendation rate of 5.3video/s. The best performance of
the proposed model is due to its consideration of the
personalized service of users and to merging the attributes of
learners and learning resources to create personalized
recommendations.
3.1.4. Analysis of a Real-Life Example.
The six sets of real data used in this paper are arranged in the
confusion matrix depicted in Fig. 8. The columns of the
matrix represent the human models generated by the
proposed model, while the rows of the matrix represent the
real labels. It can be seen that the accuracy of producing
human structural models for the six testers of the six sets of
experiments was 92.51%, 93.32%, 93.34%, and 93.62%,
respectively. These results prove that the proposed model
provides stable performance on multiple data sets besides
providing better real-time performance and hence it provides
better robustness.
Fig. 8. Confusion matrix.
In this research work, a proposed Artificial Intelligence
algorithm is presented. The purpose of the algorithm is to
improve the learning capability of higher education students
by facilitating the quick reading, locating, and
comprehending a text section in big data and in real time
attention. While improving the cognitive capability of
students, it also emphasizes the quality and self-esteem of
students. In this paper, it is shown that the proposed
algorithm exhibits better real-time performance compared to
other standard reference models such as FCNN-CF, User-
CF, and Gram-CF recommendation models, besides
outperforming these models in many aspects.
[1] Canhoto, A.I.; Clear, F. “Artificial intelligence and machine
learning as business tools: A framework for diagnosing value
destruction potential”. Bus. Horiz., 63, 183–193, 2020.
[2] Dorça, F.A.; Lima, L.V.; Fernandes, M.A.; Lopes, C.R.
“Comparing strategies for modeling students learning styles
through reinforcement learning in adaptive and intelligent
educational systems: An experimental analysis”, Expert Syst.
Appl., 40, 2092–2101, 2013.
[3] Loftus, M.; Madden, M.G. “A pedagogy of data and Artificial
Intelligence for student subjectification”, Teach. High.
Educ.,25,456–475, 2020.
[4] Alkhatlan A, Kalita J, “Intelligent tutoring systems: a
comprehensive historical survey with recent developments”,
Int J Comput Appl 975:8887, 2018.
[5] Xu, J.; Moon, K.H.; van der Schaar, M. “A Machine Learning
Approach for Tracking and Predicting Student Performance in
Degree Programs”. IEEE J. Sel. Top. Signal Process., 11,
742–753, 2017.
[6] Khare K, Stewart B, Khare A, “Artificial intelligence and the
student experience: an institutional perspective”, IAFOR J
Educ 6(3):63–78, 2018.
[7] S. Makridakis, “The forthcoming artificial intelligence (AI)
revolution: its impact on society and firms,” Futures, vol. 90,
no. jun., pp. 46–60, 2017.
[8] Abdi, S., Khosravi, H., & Sadiq, S. “Modelling learners in
crowdsourcing educational systems”, in: International
Conference on ArtificialIntelligence in Education (pp. 3–9).
Springer, 2020.
[9] A. I. Review, “About the authors,” Artificial Intelligence
Review, vol. 15, no. 6, pp. 1–6, 2016.
[10] Chounta, I. A., Bardone, E., Raudsep, A., & Pedaste, M.
“Exploring teachers’ perceptions of artificial intelligence as a
tool to support their practice in Estonian k- 12 education”,
International Journal of Artificial Intelligence in Education, 1–
31, 2021.
[11] Yigitcanlar T., Mehmood R., and Corchado J. M., “Green
Artificial Intelligence: Towards an Efficient, Sustainable and
Equitable Technology for Smart Cities and Futures,”
Sustainability, vol.13, no.16, pp:1-14, August 2021.
[12] J. P. Davis and W. A. Price, “Deep learning for teaching
university physics to computers,” American Journal of
Physics, vol. 85, no. 4, pp. 311-312, 2017.
[13] Belgaum M. R., Alansari Z., Musa S., Alam M. M., and
Mazliham M. S., “Role of artificial intelligence in cloud
computing, IoT and SDN: Reliability and scalability issues,”
International Journal of Electrical and Computer Engineering,
vol.11, no.5, pp:4458-4470, October 2021.
4. Conclusions
References
WSEAS TRANSACTIONS on COMPUTERS
DOI: 10.37394/23205.2023.22.2
Amin Al Ka’Bi
E-ISSN: 2224-2872
12
Volume 22, 2023
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
The author contributed in the present research, at all
stages from the formulation of the problem to the
final findings and solution.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The author has no conflict of interest to declare that
is relevant to the content of this article.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
