Received 21 June 2023, accepted 6 July 2023, date of publication 12 July 2023, date of current version 20 July 2023.

Digital Object Identifier 10.1109/ACCESS.2023.3294700

Prediction of Student’s Performance With

Learning Coefficients Using Regression
Based Machine Learning Models
PALLAVI ASTHANA 1 , (Member, IEEE), SUMITA MISHRA 1 , (Senior Member, IEEE),
NISHU GUPTA 2 , (Senior Member, IEEE), MOHAMMAD DERAWI 2 , AND ANIL KUMAR 1
1 Amity School of Engineering and Technology, Amity University, Lucknow Campus, Uttar Pradesh 226028, India
2 Department of Electronic Systems, Faculty of Information Technology and Electrical Engineering, Norwegian University of Science and Technology, 2815

Gjøvik, Norway
Corresponding author: Nishu Gupta ([email protected])

ABSTRACT Advanced machine learning (ML) methods can predict student’s performance with key features
based on academic, behavioral, and demographic data. Significant works have predicted the student’s
performance based on the primary and secondary data sets derived from the student’s existing data. These
works have accurately predicted student’s performance but did not provide the metrics as suggestions for
improved performance. This paper proposes the ‘Learning Coefficients’ evaluated through trajectory-based
computerized adaptive assessment. Learning coefficients also provide quantified metrics to the students to
focus more on their studies and improve their further performance. Before selecting the learning coefficients
as the key features for student’s performance prediction, their dependency on other key features is calculated
through positive Pearson’s coefficient correlation. Further, the paper presents comparative analysis of the
performance of regression-based ML models such as decision trees, random forest, support vector regression,
linear regression and artificial neural networks on the same dataset. Results show that linear regression
obtained the highest accuracy of 97% when compared to other models.

INDEX TERMS Adaptive assessment, learning coefficients, machine learning models, regression based
prediction, student’s grade prediction.

I. INTRODUCTION along with meaningful feedback to modify the environment

Educational institutes record student data as demographic, in order to improve the learning in students [2]. For vari-
academic, and educational. Many programs conducted as part ous reasons like dropout rate, retention rate and assignment
of teaching-learning activities also generate different data like performance, most of the works have considered the metrics
study behaviors, study patterns, and participation in extracur- of student’s performance as the acquired Cumulative Grade
ricular activities. This vast amount of data can efficiently Point Average (CGPA) and course grade range at the end
predict student’s performance and establish a correlation of the engineering program. Cumulative grades constitute a
between it’s features through machine learning (ML) based significant factor in determining a student’s eligibility in var-
algorithms when processed with proper tools [1]. These cor- ious aspects of progression such as participation in placement
relations can predict student’s performance at the end of the activities or applying for higher studies, etc. [3]. Consider-
program. Prediction of student’s performance is significant able research has been done in the past to predict student’s
in adapting the learning environment through customized performance due to two reasons; the availability of relevant
academic assistance, academic guidance, advice mentoring, information at educational institutes in digital format, and the
examining efficiency and effectiveness of learning methods development of ML tools that can model this information
into data to make meaningful patterns through data mining
The associate editor coordinating the review of this manuscript and and classification techniques [4]. By early identification and
approving it for publication was Vicente Alarcon-Aquino . proper intervention, teachers and institutions can provide the
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
72732 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 11, 2023
P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

necessary support for a better understanding of courses [5] Machine (SVM), Naïve Bayes, K-Nearest Neighbour (KNN),
to improve student’s knowledge and academic performance. Decision Tree (DT), Random Forests (RF), Multilayer Per-
ML models help discover patterns and relationships between ception (MP), Linear Regression (LR), bagging and boosting
data variables and analyze complex non-linear relationships that have provided accurate results while predicting perfor-
for decision-making. They are also valuable in predicting stu- mances [6], [7]. This data was pre-processed with supervised
dent’s performance based on various factors including their learning methods as shown in Figure 1.
earlier and in-term performance in each course.
B. ORGANIZATION
A. CONTRIBUTION Section II presents a background of the ML-based student’s
Earlier work relied on static predictors that only predicted the performance prediction models, discusses the feature selec-
performance without providing the measures of improvement tion process for training and testing of these models and iden-
in the performance. This works used ‘Learning Coefficients’ tifies the non-uniformity among these features. Section III
as the dynamic predictors of student’s performance. These discusses the proposed method and introduces ‘learning coef-
are academic predictors and are calculated through adap- ficients’ as quantitative metrics of student’s performance
tive assessment conducted as continuous evaluation during calculated through adaptive assessment. In section IV, learn-
the semester. They bring uniformity in one of the predic- ing coefficients are established as one of the key features for
tors, purely as the evaluation of student’s knowledge with prediction which is confirmed by calculating its correlation
respect to the course studied in the current academic semester. with other academic and demographic features. Further, this
This way it provides a true measure of student’s current section presents the results and discusses the performance
performance and better performance in adaptive assessment evaluation of various models. Finally, section V concludes the
suggests that they have a possibility of good performance in article.
their final examination also.
A correlation feature selection approach is applied in this II. BACKGROUND
work to determine which features are most important for pre- This section discusses the prominent ML models used for
dicting student’s performance. Correlation feature selection student’s performance prediction in recent research works.
scores are presented using which the features were scored Table 1 summarizes the details of such works discussing
using this method. Besides this, the major contributions of ML-based models, feature selection and the adopted method-
this work are: ology based on the size of the dataset or the primary or
(1) To identify the ML techniques prominently adopted for secondary data selection. The sets of predictors is not the
accurately predicting student’s academic performance. same in each work, bringing non-uniformity in the prediction
(2) To understand the feature selection in ML based models. results.
(3) To propose quantified features as a uniform feature for
the measurement of improvement in student’s perfor- A. FEATURES FOR PREDICTION MODELS
mance based on research gaps. Many factors like academic performance, demographic
Apart from these contributions, the article reviews many background, study behavior, selection of the courses,
standard ML based algorithms such as Support Vector and extracurricular activities affect the performance of

FIGURE 1. Prediction of student’s performance based on ML models.

VOLUME 11, 2023 72733

 P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

TABLE 1. Prominent ML models used for student’s performance prediction.

engineering students. When any model is developed based and motivational levels. Table 1 shows that most of the work
on this data, it provides ideal conditions which may not exist focus on implementing and validating the effectiveness of
for many students as they have different kinds of background the predictive algorithms. Researchers have selected different

72734 VOLUME 11, 2023

P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

features in their work for inclusiveness and to increase the different backgrounds, study habits and variable academic
range of features. Most of the features are static in nature and benchmarks. These factors can predict the student perfor-
suffer from many limitations such as: mance but as these features are not sufficient to record the
student’s learning during the semester, there is a need to
(i) Academic data: the most reliable data to predict stu-
devise an effective tool to predict students’ performance
dent’s academic performance. It consists of student’s
before the final assessment that may improve learning out-
CGPA, internal assessment, internal examination scores
comes. Therefore, there is a need to devise an effective tool
and courses selected by the students. Major problem
to predict and improve students’ performance. In this work,
with this kind of data is non-uniformity. Different insti-
we aim to predict their performance during the semester to
tutes have their own assessment methods and any model
provide with the opportunities of improvement [26]. This can
that has shown accuracy on a dataset of any institute
be done with the inclusion of a dynamic feature along with
may not provide the same results for another institute.
other features to ensure the measured improvement.
Moreover, the applied methods may differ for differ-
ent courses. For example, with the secondary school
data, few institutes have a higher cut-off for admission. III. PROPOSED METHOD
Hence, there will be less variation in this data, whereas, To strengthen the predictive models, ‘learning coefficients’
in the institutes with a lower cut-off, this variation will calculated through adaptive assessment [28] are introduced
be high. Therefore, the reliability of the same features in this work. They are the quantified performance metrics for
would be different in both the settings [25], [26]. a group of courses including a course and its pre-requisite
(ii) Demographic and personality data: such details include courses. It has been established that performance in the
the age, gender and personality data about student’s pre-requisite course is a significant predictor of performance
background like parent’s education, parent’s income, in the successive courses [18]. This academic feature depends
emotional intelligence, student’s interests, level of moti- on the knowledge-building of the students thereby bringing
vation, communication, interest in sports, hobbies and uniformity in the dataset. In addition, adaptive assessment
ethnicity. This data is vital because student’s motiva- can be conducted multiple times during a semester, thus
tional level heavily rely on these factors. However, providing more opportunities for improvement. To assess
accurate data is often missing, and analysis with such the performance of learning coefficients based student’s per-
kind of data can lead to biases against students with formance prediction model, a study was conducted on the
a specific background. Also, this kind of data would students of undergraduate Computer Science and Engineer-
require permission concerning ethics [27]. ing (CSE) program at Amity University, Lucknow, India.
(iii) Institutional data: it pertains to the facilities depen- The conceptual framework of the proposed methodology is
dent on the institutes, the condition of the laborato- shown in figure 2. The university follows Choice Based
ries, the state of experiments, infrastructure, teaching Credit System (CBCS). All the students need to complete
methodology, transportation facility and communica- a minimum number of credit units in each semester. The
tion medium [26], [27]. This data is contextual and program structure defines the courses in each semester under
dependent on factors such as the availability of proper four different categories namely, Core Courses (CC), Domain
resources for selected courses that would affect the accu- Elective (DE), Value Added Courses (VAC) and Open Elec-
racy of the predictive models. tive (OE). CC and DE include courses that are related to the
(iv) Behavioral data: it includes the study pattern, attention branch a student is enrolled in, while OE and VAC allow the
span, rate of downloading of study material in case of students to study topics of interest that are beyond their core
flipped classes, social interactions, time spent on social domain. DE courses are taught in collaboration with other
sites and playing educational computer games. Feature- engineering departments. CSE program has domain elective
related behavior of students is crucial for the analysis of course on Microprocessors (MP) in the seventh semester
their performance prediction. Most of this data is subjec- which is offered by electronics and communication engineer-
tive, based on surveys and questionnaires. Researchers ing department. This experiment is conducted on 91 students
use this as primary data or extract secondary information who opted MP as DE course. Two prerequisite courses
from this data for prediction purposes. This data relies for the MP course as mentioned in the program structure
heavily on the student’s responses which can have many are (a) Computer Organization and Architecture (COA) and
diversification and needs to precisely map various fea- (b) Basics of Electronics Engineering (BEE). Students study
tures like the demographic and academic profile of the these prerequisite courses in the third and fourth semester
students [26], [27]. respectively.

B. GAPS IN THE PRESENT RESEARCH A. TRAJECTORY BASED ADAPTIVE ASSESSMENT

The main features utilized for the student’s performance Adaptive assessments are tailored specifically to each stu-
prediction are demographic details, institutional details and dent by providing a customized set of questions where the
educational details which are static in nature. Students have difficulty level of every succeeding question depends on

VOLUME 11, 2023 72735

 P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

of appearance: BEE → COA→ MP. A correct answer will

award score 1, otherwise 0. The assessment is formative. The
process of assessment of scores is shown in figure 4.
(i) The first question in the set is from the course BEE; if a
student answers it correctly, the score is 1. If it is wrong,
the score is 0. This score is ‘a’. Cumulative a’s are ‘J’.
(ii) If a=1, the next question will appear from the course
COA; if a=0, the next question is from BEE again. For
a correct answer, the score is 1 otherwise 0. This score
is ’b’. Cumulative b’s are K.
FIGURE 2. Adopted framework. (iii) If b=1, the next question will appear from the course
MP; if b=0, the next question is from COA again. For a
correct answer, the score is 1, otherwise 0. This score is
c Cumulative c’s are L.
(iv) Either c=0 or c=1, the next question will appear from
BEE for second set of questions.
(v) This process will end when L adds ten scores. Learning
coefficients J, K, and L are the average values of ai , bi
and ci respectively as shown in the equations (1), (2) and
(3) respectively.

J = (a1 + a2 + . . . . . . ..a10 )/10 (1)

K = (b1 + b2 + . . . . . . b10 )/10 (2)
L = (c1 + c2 + . . . . . . ..c10 )/10 (3)

where ai → Score in the questions of course BEE bi → Score

in the questions of course COA ci → Score in the questions
of course MP. Scores are calculated for all the students to
compare the performance in each course in the trajectory.
Evaluated learning coefficients are shown in figure 5. Aver-
FIGURE 3. Designing of a computerized assessment.
age values of the learning coefficients for 91 students are
calculated. Average value of J is 0.94, K is 3.3 and L is 6.2.
These are the metrics for the measurement of the academic
the student’s performance in the preceding question. The knowledge when taught by trajectory based teaching peda-
trajectory-based adaptive assessment works in this pattern. gogy, so these are purely academic features for the prediction.
Automated method of adaptive question generation is still
in the design phase but its basic building blocks are shown IV. RESULTS AND DISCUSSIONS
in figure 3. This system will require the blended learning for Data analysis was conducted in two phases (i) Establish-
the implementation, where questions will be uploaded on web ment of ‘learning coefficients’ as key features for prediction
server and accessed through Web browser. by finding the correlation with academic and demographic
features and (ii) Prediction of student’s CGPA based on the
B. PARTICIPANTS AND DATABASE academic features and learning coefficients.
This study was conducted on 91 undergraduate students of Experiments ran on open development called collaboratory
CSE department. Their records were grades in Microproces- (colab) developed by Google for analysis, and supports full
sor course, their SGPA for seven semesters, cumulative grade Python syntax. It is an efficient method to run ML-based
pointers at the course completion, and board of secondary algorithms. Colab provides the facility to plot graphs that
education. The student’s names and enrolment numbers were facilitate the graphical representation and is easier to com-
removed to maintain the privacy of data and for the analysis prehend. Results are saved on Google drive directly. Data
within the university’s data privacy obligations. Pre-processing: Microsoft excel file is saved in .csv format
for processing colab. Critical stages of data pre-processing
C. CALCULATION OF LEARNING COEFFICIENTS are: Dataset Cleaning: removing features such as student’s
AS EVALUATION METRIC names, enrolment number, course descriptions, batch, insti-
In order to calculate the learning coefficients, we present an tution and date of birth. University maintains the records of
algorithm. The question paper consists of 10 sets of multiple all students on an online portal. Hence, no data was missing
choice-based options, each set of three questions in the order and all the records were complete.

72736 VOLUME 11, 2023

P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

FIGURE 4. Calculation of scores in adaptive assessment.

FIGURE 6. Methodology for the evaluation of the prediction model.

FIGURE 5. Values of ‘learning coefficients’ J, K and L. pointers in class 10th and 12th , SGPA of semester I to VII,
marks obtained in the end semester exam of the Micropro-
cessor course, and gender as a demographic feature. The
dependency between the learning coefficients and response
A. FEATURE ENCODING
variable was calculated through a correlation matrix as shown
All the data types are converted to numeric data type because in figure 7. Correlation coefficients are calculated through
the algorithm used for this work supports numeric data type Scikit, a software machine learning library used for Python
as shown in Table 2. Grade pointers and scores were not programming language. The colab tool was used for the
rounded-off to integers but taken as original values for the evaluation of the work that supports Scikit which further
prediction with values up to two places after the decimal. calculates the correlation coefficient by Pearson’s correlation
Few features such as gender and board of examination (high coefficient (r). The correlation coefficient is calculated as
school, intermediate) were converted to numeric values mak- shown in equation 4:
ing it convenient to run on Python.
6(xi − x)(yi − y)
r=p (4)
B. FEATURES SELECTION AND DATA PRE-PROCESSING 6(xi − x)2 6(yi − y)2
Before the evaluation phase, pre-processing included remov- Correlation matrix shows the correlation between learn-
ing extra columns from dataset and converting them into ing coefficients and other key features. Positive correlation
numerical values to make it usable for the analysis. Figure 6 coefficient value (r) shows the dependency among features
shows the methodology considered for the evaluation of the and a negative correlation value indicates independence in
prediction model. features. The correlation coefficient was calculated for learn-
ing coefficients and the remaining features. Calculated values
C. CORRELATION BETWEEN LEARNING COEFFICIENTS of ‘r’ are favorable for the grade pointers in the course
AND OTHER KEY FEATURES Microprocessor ranging between .70 to .78; CGPA of all the
Learning coefficients are correlated with academic features seven semesters between .23 to .82; CGPA achieved in final
and other demographic features for establishing them as pre- semester as .73 to .84; percentage of 10th as .31 and .36 in
dictor of performance. Academic data consists of student’s the 12th examination. Positive correlation coefficient values

VOLUME 11, 2023 72737

 P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

TABLE 2. Selection of key attributes.

1) Linear Regression
LR is used to quantify the linear relationship between
an explanatory variable and response variable. In the
event, if there are more than one explanatory vari-
ables, then it is called Multiple Linear Regression. This
kind of relationship predicts the response variable or
dependent variable, and the variables for predicting the
variable are called the explanatory or independent vari-
ables [29], [30]. The difference between the true and
predicted values is known as the residual. Predicting
a response variable based on the explanatory variable
is known as regression analysis, and a straight line
that describes the prediction is known as the regression
line. The calculation of the best fit line is based on
the method of least squares which minimises the sum
of the vertical distance between all of the data points
and the line of best fit. Mathematical representation of
multiple linear regression is given in equation 5.
FIGURE 7. Correlation matrix.
yi = α0 + α1 xi1 + α2 xi2 + . . . ..αm xim + ϵi (5)
where xi denotes explanatory variable, yi is the
prove that learning coefficients are also significant factors in response variable, α0 , α1 . . . .αn are model parameters
predicting the student’s performance. wherein α0 is also known as the intercept and ϵi is
the residual. The subscript ‘i’ refers to the iþ data
D. EVALUATION OF THE MACHINE LEARNING MODELS instances in the dataset and ‘m’ refers to the number
FOR PREDICTION of explanatory variables.
The dataset was split into training and testing data in 90% 2) Decision Tree
and 10% respectively of the total inputs. As the dataset DTs classify data into good performing and poor
is small, regression-based ML models are most suitable to performing based on the relationship between the fea-
run the prediction. Deep Learning based models are prone tures and their comparative significance. DT regression
to over-fitting on a small dataset. After pre-processing, the trains a model in the form of a tree to predict data in the
dataset is run for the following ML models: LR, DT, RF, future [31]. They are simple algorithms for prediction
and SVR on the colab tool which are briefly discussed structured as root to a leaf node in a tree. Node repre-
below. sents the test of features indicating their possible value

72738 VOLUME 11, 2023

P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

FIGURE 9. Comparison between the true values and predicted values

in LR.
FIGURE 8. Support Vector Regression: hyperplane line with boundary
epsilon.

model predicts values that are either lower or higher than the
actual value. Therefore, we determine the model’s accuracy
and creates the split based on the calculated value of
through evaluation metrics based on residuals. Residuals are
the feature of the decision. This split creates new nodes
the difference between the actual and the predicted values.
and the process continues until the formation of pure
Important parameters for the models such as accuracy RMSE,
leaf nodes.
MAE, and MSE [33], [34] have been evaluated. Evaluation
3) Random Forests
metric used to measure accuracy of best performing model
Random forests have many DTs that take different
is R2 score. It is a simple but efficient metric to compare the
features as a root and then based on it Entropy and GINI
developed models. R2 score is calculated using the formula
Index are calculated. Each tree may have a different
as in equation 6.
decision. An algorithm based on RF selects the decision
chosen by most DTs based on the majority voting RSS
R2 = 1 − (6)
system [18]. To apply RF algorithm, TSS
• M number of data instances are chosen from the
where R2 is the coefficient of determination, RSS is the sum
dataset. of the squares of the residuals and TSS is the total sum of
• A DT associated to these M data instances is then P
2
the squares. Here, RSS = yi − ŷi where yi is the actual
built.
(yi − ȳ)2
P
value and ŷi is the predicted value; TSS =
• Both the steps listed above are repeated until N
where yi is the actual value and ȳ is the mean value of the
number of DTs are generated.
variable/feature.
For a new input data point, each tree generates the In order to determine various parameters, ML models were
prediction value of response variable and assigns that developed using the default value settings of most of the
data point to the average across all the predicted values. hyper-parameters defined during the experiment and are pre-
4) Support Vector Regression (SVR) sented in Table 3.
SVR is used to model non-linear relationships between
variables to adjust the model’s robustness through esti- TABLE 3. Parameter settings for ML algorithms.
mated hyperplane functions. A hyperplane is the best
fit line with maximum points fitted within a threshold
value. They are the decision boundaries to predict the
continuous output based on a set of mathematical func-
tions known as kernels. Popular kernels used in SVR
are linear, non-linear, polynomial, radial basis function
(RBF) and sigmoid [32]. Figure 8 shows the hyperplane
between two variables X1 X2 . ϵ is a tunable parame-
ter that determines the width of the plane around the
hyperplane. Points that fall inside this plane are correct
predictions. SVR is reasonable for small datasets as it To assess the performance of ML models, apart from R2
has good generalization capability. Score, MAE (Mean Absolute Error), MSE (Mean Square
It is not possible for a regression-based model to pre- Error), and RSME (Root Mean Square Error) are also
dict the exact value of a continuous variable. A regression evaluated.

VOLUME 11, 2023 72739

 P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

RMSE, which is the standard deviation of the differences

between the predicted values p′i and observed (true) values p′i ,
is calculated in equation 7.
sP
n ′ 2
i=1 (pi − pi )
RMSE = (7)
n
MAE is the magnitude of the difference between the pre-
diction of observation and the actual value of observation.
MAE is the average of the errors in the entire group, used as a
loss function for regression-based ML algorithms to provide
quantifiable measure of the errors as shown in equation 8.
Pn
p′ − pi FIGURE 10. Training loss for neural network.
MAE = i=1 i (8)
n
MSE is the loss function calculated as squared errors over
the entire group as shown in equation 9. The main advantage errors due to over-fitting. maximum tree depth for DT model
of calculating MSE is to ensure that there are no outliers in was initially kept at 3 then it was increased to 5.
the model. Figure 9 shows the comparison between true values and
Pn 2
predicted values plotted o test data. Predicted values are
p′ − pi not deviated much from the true values, thus giving better
MSE = i=1 i (9)
n accuracy.
Prediction models based on LR, RF, DT and SVR were Apart from the models shown in Table 4, the given dataset
trained and tested on the Scikit tool that supports Python. also ran on the Artificial Neural Network, which is also a
DT and RF are also regression models of ML with a similar regression-based model. For this work, an ANN was trained
loss function as MSE and MAE. with four layers one input layer, two hidden layers, and one
output layer. The mean square error is the loss function with
TABLE 4. Calculated values of Accuracy RMSE, MAE and MSE. the optimizer ‘adam’ and ‘ReLU’ activation[36]. The model
ran for 100 epochs and provided an accuracy of 79% with a
mean absolute error of .05, as shown in figure 11.

Table 4 shows the calculated evaluation metrics from dif-

ferent models, and it shows that LR outperforms the other
models for the test data. It is expected as complex models
perform well on training data but due to over-fitting they,
perform poorly over the test data. The table also shows the
calculated values of accuracy of RMSE, MAE, and MSE for
the above LR, RF, DT and SVR algorithms. LR provides the
maximum accuracy of 97%, values of the residuals are least,
and hence the values of errors are also small as .0149 .0183,
and .0003 for RMSE, MAE and MSE respectively. As we
can see from the formula, MAE measures the average magni- FIGURE 11. Comparison between true values and predicted values in RF.
tude of the errors without considering their direction. It’s the
average of the absolute differences between predictions and These parameters show that along with the traditional pri-
actual values over the test sample. Apart from determining mary or secondary feature, the proposed learning coefficient
how close the prediction is to the actual value on an average, is also an effective feature that can predict the student’s
RMSE also indicates the effect of large errors. Scikit-learn performance. Along with prediction, it is also a quantified
evaluation metric library used in this work does not have measure that suggests improvement metrics. Existing works
RMSE function. So, to get the RMSE, we use the Numpy focus on predicting the performance of students, without
square root method to find the square root of mean squared providing evidential measures of improvement [35]. One of
error. RMSE function values are examined to determine if the significant limitations of our work is small dataset having
there are any large errors in the model developed. We can only 91 records. Research with more data may result in
see from table 4 that the RMSE value is larger than the MAE, better feature engineering and noteworthy insights. Proposed
particularly in the DT model. This is a result of some large method provides a valuable tool for course instructors to

72740 VOLUME 11, 2023

P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

modify teaching practices for imparting quality education. [7] J. Xu, K. H. Moon, and M. van der Schaar, ‘‘A machine learning approach
for instance, there may be provision of additional teaching for tracking and predicting student performance in degree programs,’’
IEEE J. Sel. Topics Signal Process., vol. 11, no. 5, pp. 742–753, Aug. 2017.
support to poor performing students at early stage. [8] M. I. Hoque, A. K. Azad, M. A. H. Tuhin, and Z. U. Salehin, ‘‘University
students result analysis and prediction system by decision tree algorithm,’’
Adv. Sci., Technol. Eng. Syst. J., vol. 5, no. 3, pp. 115–122, 2020.
V. CONCLUSION
[9] B. Khan, M. S. H. Khiyal, and M. Daud Khattak, ‘‘Final grade prediction
The work proposed in this article utilized regression based of secondary school student using decision tree,’’ Int. J. Comput. Appl.,
ML models for student’s performance prediction. Out of the vol. 115, no. 21, pp. 32–36, Apr. 2015.
various models that have been evaluated, 97% prediction [10] T. Jeevalatha, N. A. N. Ananthi, and D. S. Kumar, ‘‘Performance analysis
of undergraduate students placement selection using decision tree algo-
accuracy has been achieved by linear regression model. Data rithms,’’ Int. J. Comput. Appl., vol. 108, no. 15, pp. 27–31, Dec. 2014.
sets consist of the academic data including student’s CGPA [11] A. Polyzou and G. Karypis, ‘‘Feature extraction for classifying students
for seventh semester, grades pointers in class 10th and 12th based on their academic performance,’’ in Proc. 11th Int. Conf. Educ. Data
Mining, Buffalo, NY, USA, Jul. 2018, pp. 356–362.
examinations, and gender of the students. We employed a [12] A. Acharya and D. Sinha, ‘‘Early prediction of students performance
novel set of variables namely learning coefficients for deter- using machine learning techniques,’’ Int. J. Comput. Appl., vol. 107, no. 1,
mining the student performance, which have not been used pp. 37–43, Dec. 2014.
[13] T. Anderson and R. Anderson, ‘‘Applications of machine learning to
in prior works. The learning coefficients were calculated for student grade prediction in quantitative business courses,’’ Glob. J. Bus.
the course of Microprocessor taught to the seventh-semester Pedagog., vol. 1, no. 3, pp. 13–22, 2017.
students, calculated in trajectory based adaptive assessment [14] M. A. Al-Barrak and M. Al-Razgan, ‘‘Predicting students final GPA using
decision trees: A case study,’’ Int. J. Inf. Educ. Technol., vol. 6, no. 7,
of three inter-related courses. Traditional key features are pp. 528–533, 2016.
based on the student’s academic data scored at the end-term [15] L. M. A. Zohair, ‘‘Prediction of student’s performance by modelling small
examination. However, the learning coefficients are evaluated dataset size,’’ Int. J. Educ. Technol. Higher Educ., vol. 16, no. 1, pp. 1–18,
Dec. 2019.
during the continuous evaluation conducted throughout the [16] L. Cagliero, L. Canale, L. Farinetti, E. Baralis, and E. Venuto, ‘‘Predicting
semester. Hence, their inclusion as the key features provides a student academic performance by means of associative classification,’’
quantitative metric to the students so that they can focus more Appl. Sci., vol. 11, no. 4, p. 1420, Feb. 2021.
[17] A. Siddique, A. Jan, F. Majeed, A. I. Qahmash, N. N. Quadri, and
on their studies and secure better grades in the end semester M. O. A. Wahab, ‘‘Predicting academic performance using an efficient
examination. Whereas larger dataset would certainly be a bet- model based on fusion of classifiers,’’ Appl. Sci., vol. 11, no. 24, p. 11845,
ter choice, future extension of this work will include a bigger Dec. 2021.
[18] W. Wunnasri, P. Musikawan, and C. So-In, ‘‘A two-phase ensemble-based
sample size to obtain the improved overall performance of method for predicting learners’ grade in MOOCs,’’ Appl. Sci., vol. 13,
the proposed ML models. This work is still in the research no. 3, p. 1492, Jan. 2023.
stage and proposed ‘Learning Coefficients’ are the innovative [19] A. Zollanvari, R. C. Kizilirmak, Y. H. Kho, and D. HernáNdez-Torrano,
‘‘Predicting students’ GPA and developing intervention strategies
evaluation metrices of student’s performance. Purpose of this based on self-regulatory learning behaviors,’’ IEEE Access, vol. 5,
work is to corelate the learning coefficient with other aca- pp. 23792–23802, 2017.
demic predictors and use it to predict student’s performance. [20] T. M. Alam, M. Mushtaq, K. Shaukat, I. A. Hameed, M. U. Sarwar, and
S. Luo, ‘‘A novel method for performance measurement of public edu-
This work was conducted in only one course that limited cational institutions using machine learning models,’’ Appl. Sci., vol. 11,
the data. For the less data, simpler ML based model such no. 19, p. 9296, Oct. 2021.
as Linear regression has provided accurate results. But as, [21] S. Sakri and A. S. Alluhaidan, ‘‘RHEM: A robust hybrid ensemble model
for students’ performance assessment on cloud computing course,’’ Int.
the learning coefficient will be calculated for more courses J. Adv. Comput. Sci. Appl., vol. 11, no. 11, pp. 388–396, 2020.
and programs, it will enlarge the dataset, and more com- [22] A. Polyzou and G. Karypis, ‘‘Feature extraction for next-term prediction
plex Machine learning based models will be required for the of poor student performance,’’ IEEE Trans. Learn. Technol., vol. 12, no. 2,
pp. 237–248, Apr. 2019.
prediction. [23] R. Hasan, S. Palaniappan, S. Mahmood, A. Abbas, K. U. Sarker, and
M. U. Sattar, ‘‘Predicting student performance in higher educational insti-
tutions using video learning analytics and data mining techniques,’’ Appl.
REFERENCES
Sci., vol. 10, no. 11, p. 3894, Jun. 2020.
[1] B. Albreiki, N. Zaki, and H. Alashwal, ‘‘A systematic literature review [24] Y. Baashar, Y. Hamed, G. Alkawsi, L. F. Capretz, H. Alhussian,
of student’ performance prediction using machine learning techniques,’’ A. Alwadain, and R. Al-amri, ‘‘Evaluation of postgraduate academic per-
Educ. Sci., vol. 11, no. 9, p. 552, Sep. 2021. formance using artificial intelligence models,’’ Alexandria Eng. J., vol. 61,
[2] C. Romero and S. Ventura, ‘‘Guest editorial: Special issue on early pre- no. 12, pp. 9867–9878, Dec. 2022.
diction and supporting of learning performance,’’ IEEE Trans. Learn. [25] A. Hellas, P. Ihantola, A. Petersen, V. V. Ajanovski, M. Gutica,
Technol., vol. 12, no. 2, pp. 145–147, Apr. 2019. T. Hynninen, A. Knutas, J. Leinonen, C. Messom, and S. N. Liao, ‘‘Pre-
[3] A. Khan and S. K. Ghosh, ‘‘Student performance analysis and prediction dicting academic performance: A systematic literature review,’’ in Proc.
in classroom learning: A review of educational data mining studies,’’ Educ. Companion 23rd Annu. ACM Conf. Innov. Technol. Comput. Sci. Educ.,
Inf. Technol., vol. 26, no. 1, pp. 205–240, Jan. 2021. Jul. 2018, pp. 175–199.
[4] M. Al-kmali, H. Mugahed, W. Boulila, M. Al-Sarem, and A. Abuhamdah, [26] M. Bilal, M. Omar, W. Anwar, R. H. Bokhari, and G. S. Choi, ‘‘The role of
‘‘A machine-learning based approach to support academic decision- demographic and academic features in a student performance prediction,’’
making at higher educational institutions,’’ in Proc. Int. Symp. Netw., Sci. Rep., vol. 12, no. 1, p. 12508, Jul. 2022.
Comput. Commun. (ISNCC), Oct. 2020, pp. 1–5. [27] F. Qiu, G. Zhang, X. Sheng, L. Jiang, L. Zhu, Q. Xiang, B. Jiang, and
[5] F. Ofori, E. Maina, and R. Gitonga, ‘‘Using machine learning algorithms to P.-K. Chen, ‘‘Predicting students’ performance in e-learning using learning
predict students’ performance and improve learning outcome: A literature process and behaviour data,’’ Sci. Rep., vol. 12, no. 1, p. 453, Jan. 2022.
based review,’’ J. Inf. Technol., vol. 4, no. 1, pp. 33–55, 2020. [28] P. Asthana, S. Tanwar, A. Kumar, and S. Mishra, ‘‘Students’ assessment
[6] B. Sekeroglu, K. Dimililer, and K. Tuncal, ‘‘Student performance predic- for quantitative measurement of course learning outcomes in online class
tion and classification using machine learning algorithms,’’ in Proc. 8th of power plant instrumentation,’’ in Proc. Int. Conf. Advancement Technol.
Int. Conf. Educ. Inf. Technol., Mar. 2019, pp. 7–11. (ICONAT), Jan. 2022, pp. 1–5.

VOLUME 11, 2023 72741

 P. Asthana et al.: Prediction of Student’s Performance With Learning Coefficients

[29] M. Gadhavi and C. Patel, ‘‘Student final grade prediction based on linear NISHU GUPTA (Senior Member, IEEE) received
regression,’’ Indian J. Comput. Sci. Eng., vol. 8, no. 3, pp. 274–279, 2017. the Ph.D. degree in electronics and communica-
[30] B. Sravani and M. M. Bala, ‘‘Prediction of student performance using tion engineering from MNNIT, Allahabad, India,
linear regression,’’ in Proc. Int. Conf. Emerg. Technol. (INCET), Jun. 2020, in 2016. He is currently a Postdoctoral Fel-
pp. 1–5. low (ERCIM Alain Bensoussan Fellowship) with
[31] V. Matzavela and E. Alepis, ‘‘Decision tree learning through a predictive the Department of Electronic Systems, Faculty
model for student academic performance in intelligent m-learning environ- of Information Technology and Electrical Engi-
ments,’’ Comput. Educ., Artif. Intell., vol. 2, Oct. 2021, Art. no. 100035.
neering, Norwegian University of Science and
[32] C. Yan, X. Shen, F. Guo, S. Zhao, and L. Zhang, ‘‘A novel model modifica-
Technology (NTNU), Gjøvik, Norway. He was a
tion method for support vector regression based on radial basis functions,’’
Struct. Multidisciplinary Optim., vol. 60, no. 3, pp. 983–997, Sep. 2019. member of the Zero Trust Architecture Working
[33] S. D. A. Bujang, A. Selamat, R. Ibrahim, O. Krejcar, E. Herrera-Viedma, Group of MeitY-C-DAC-STQC Project under e-Governance Standards and
H. Fujita, and N. A. Md. Ghani, ‘‘Multiclass prediction model for stu- Guidelines, Ministry of Electronics and Information Technology (MeitY),
dent grade prediction using machine learning,’’ IEEE Access, vol. 9, Government of India. His Ph.D. research was devoted to the development
pp. 95608–95621, 2021. of MAC protocols for safety applications in VANETs. He is the author and
[34] A. Namoun and A. Alshanqiti, ‘‘Predicting student performance using data the editor of several books and book chapters with publishers, including
mining and learning analytics techniques: A systematic literature review,’’ Springer, Taylor & Francis, Wiley, and Scrivener Publishing. His research
Appl. Sci., vol. 11, no. 1, p. 237, Dec. 2020. interests include intelligent vehicles, the IoT, smart city and transportation,
[35] Y. Meier, J. Xu, O. Atan, and M. van der Schaar, ‘‘Predicting grades,’’ IEEE and augmented cyber-security. He was a recipient of the Best Paper Presen-
Trans. Signal Process., vol. 64, no. 4, pp. 959–972, Feb. 2016. tation Award from NTU, Singapore, in 2019. He serves as a reviewer for
various high quality journals.

PALLAVI ASTHANA (Member, IEEE) is currently

working as a full time Assistant Professor in the
Department of Electronics and Communication
Engineering at Amity University, Uttar Pradesh,
India. She has done her M.Tech. in Digital Systems
and is pursuing Ph.D. in Computer Science and MOHAMMAD DERAWI received the Diploma,
Engineering. Her area of interest includes Instru- B.Sc., and M.Sc. degrees in computer science
mentation, machine learning, Control Systems etc. engineering from the Technical University of
Currently, she is working in the field of Engineer- Denmark, in 2007 and 2009, respectively, and
ing Education and developing the tools for the the Ph.D. degree in information security from
quantitative measurement of Course learning Outcomes. She has published the Norwegian Information Security Laboratory
more than 25 research papers in Reputed Conferences and Journals. She has (NISLab). He is currently a Full Professor and
also written 6 book chapters. Pallavi Asthana is a member of the IEEE and received the title as the Youngest Professor of
IETE, UK. She has been a key contributor in organizing various International Norway within Electronic Engineering and is spe-
Conference, and workshop at Amity University. She is also a member of cialized within information security, biometrics,
Internal Quality Assurance Cell at Amity University. smart devices, programming, and micro-controllers. He was active in several
EU-projects, including the 7th Framework European project TrUsted Revo-
cable Biometric IdeNtitiEs (TURBINE) and Erasmus Mundus. His research
interests include smart technologies, biometrics with specialization on both
behavioral biometrics (gait recognition and ECG) and biological biometrics
SUMITA MISHRA (Senior Member, IEEE) (fingerprint, face) in mobile devices, pattern recognition, and e-health.
received the M.Tech. degree in optical communi-
cation from the Shri Govindram Seksaria Institute
of Technology and Science, Indore, India. She
received Ph.D. degree from Dr. Ram Manohar
Lohia Avadh University, Faizabad, India. She
is currently affiliated with the Electronics and
Communication Engineering Department, Amity
School of Engineering and Technology, Amity ANIL KUMAR is currently the Deputy Pro-
University, Lucknow Campus, Uttar Pradesh, Vice Chancellor with Amity University, Lucknow,
India. Dr. Mishra is a member of IET (UK), IAENG and served as Senior India, and the Director of the Amity School of
Member of IEEE. She has more than 50 publications in international journals Engineering and Technology, Amity University.
and prestigious conferences. She has published two patents and serves as He has a vast experience of over 30 years in
reviewer and editorial board member of many highly reputed journals. the areas of modeling and simulation, navigation,
Dr. Mishra has had key responsibilities in organizing several International Kalman filtering, advanced control theory, INS,
Conferences and various international events. She has also delivered Invited extended Kalman filters, optimal estimation, flight
Speech at national and international forums. Her current research interests dynamics, and flight simulation. He has published
include deep learning, visible light communication, and Internet of things more than 40 technical articles.
and is guiding several Ph.D. students in this area.

72742 VOLUME 11, 2023