FEATURE EXTRACTION FOR CLASSIFYING STUDENTS BASED ON THEIR

ACADEMIC PERFORMANCE
Shaik Yacoob, S.S.D.Bhavani, S.Sai Sri Ram, A.Lakshmana Rao, G.Vineela
Department of Computer Science and Engineering,GIET,Rajahmundry,India

Abstract:In today's world of education, building tools to help methods. We clearly classify two groups of students: those
students learn in a formal or online way is a major challenge. The who are likely to complete a activity or career successfully
first stages of using machine learning techniques are to enable and those who seems struggling. Following the recent team
those technologies that focus on predicting student achievement in identification, we can provide additional resources and
terms of marks obtained. The downside to these methods is that support to help them succeed.
they do not work very well in predicting successful students. The On the other hand, "success" and "failure," can be related
purpose of our efforts is twofold. To begin with, we are
or not. Grade B, for example, can be a bad grade for a good
investigating whether underperforming students can be accurately
student but a good grade for a very poor student. We looked
predicted by restoring this function as a problem of binary
at many ways to divide students in the subject, such as student
separation. Second, in order to learn more about factors that may
contribute to improper functioning, we have created a set of
failure, drop-out, doing worse than predicted, and doing
human translators that measure these factors. These features are worse than expected in view of the difficulties of the course.
taken from the University of Minnesota community center grades. We have developed features or characteristics that capture
We conducted research based on these factors to identify the the various elements that affect grades at the end of the year
different groups of students who were interested in determining to gain a good understanding of the learning process and its
their value. most important qualities. We present a comprehensive
Keywords: Python, Machine Learning, Classification analysis using these features for answering following two
algorithms. questions: What factors determine a learner's success? What
are the most important factors? Outcomes are important
because different factors are very important in various
I. INTRODUCTION classification problems.

II. LITERATURE SURVEY

Higher education institutions are always looking for ways
to increase student retention and achievement. According to
1. Forecasting student performance: An educational web-
the National Center for Education Statistics in the United
based system using data mining methods: Researchers
States [8], Sixty percent of four-year undergraduate students have a unique opportunity to learn how students learn and
will not graduate from the same school where they started in what learning strategies lead to success thanks to the newly
the first six years. At the same time, 30% of first-year developed Web-based technology. Web-based systems tend
college students leave after their first year. As a result, to capture large amounts of data about user behaviour, which
universities are looking for effective and efficient ways of can be diverted using data mining techniques. This study
working for students. This is where data mining comes in to describes how to identify students and predict their final
help solve some of these difficulties. Educational data grade using attributes derived from data entered into a web-
mining and learning analytics were created to give tools to based education system. We create, evaluate, and compare
aid the learning process, such as monitoring and measuring the performance of a number of pattern separators using data
student progress, as well as predicting success and guiding from online course databases. The combination of multiple
intervention techniques. separators improves performance by highly differentiating. In
addition, we increase the accuracy of prediction by using the
The majority of present techniques are aimed at identifying
genetic algorithm (GA) to improve the accurate measurement
students who are at risk of failing to finish a course or activity
of used data. Compared with non-GA filters, GA has been
and who may benefit from additional support. One of the most shown to improve the performance accuracy of integrated
important tasks in this procedure is to forecast the student's partitions by approximately 10% to 12%. This approach can
grade achievement. While good prediction accuracy has been be very helpful in identifying vulnerable students early,
obtained, the algorithms suggested to detect low-performing especially in large classes, and allowing the teacher to
pupils have a fundamental flaw. Typically, these models are provide timely counselling.
too optimistic about student performance since the vast 2. Regression models based on knowledge for prediction
majority of students do well or well. of next-term grade:
We examine the challenge of predicting student
performance at the end of the semester before taking a course Students and their advisors can use grade-level predictions
in this paper. The issue of prediction is defined as the task of of future courses that they have not yet taken to assist them in
classification, in which two groups of students are formed selecting courses and developing appropriate qualification
based on their academic performance, to focus on plans and reviewing them according to their performance. For
underperforming students who are most in need of these the objective of next-term grade prediction, we provide a
data-based data collection model with different subject

1
knowledge gaps in this study. This approach captures the 2. Taking those records from dataset which contains
relationship between the subjects in terms of the information dropout students and assign feature value as 1 (Wgr)
they give by combining historical student-course grades with 3. Taking those records from dataset which has grades
information about the courses. lower than expected and assign feature value as 2
3. Grade prediction with temporal course-wise influence: (RelF)
4. Taking those records from dataset which has grade
New educational technology apps that evaluate data value lower than expectation and he is having difficulty
generated by universities are urgently needed to ensure that in study course and assign value as 3 (RelCF)
students graduate on time (4 to 6 years) and are properly 5. Rest student we are marking with feature value as 4
prepared for careers in their respective fields of study. In this which indicate student is performing well.
study, we provide a novel method for evaluating historical By using ‘University of Minnesota’ grade dataset we are
educational records from a big, public institution in order to extracting above features and assign those values as the target
do next-term grade prediction, or estimating the grades that a or class label for this dataset. After extracting features we are
applying 4 machine learning algorithms on this dataset to
student will receive in a course that he or she will enrol in the
generate training model, later new student record will be
next term. Accurate next-term grade prediction provides the applied on this dataset to classify that student records as good
possibility of improved student degree planning, performer or poor performer and we can know the reason of
individualized advising, and automated interventions to help poor performance such as Fgr (indicate as failing student),
students stay on track in their chosen degree programme and Wgr (indicate as dropout), ReIF (lower than expected grade)
graduate on time. We introduce Matrix Factorization with or RelCF (lower than expected grade and having course
Temporal Course-wise Influence, a factorization-based difficulty).
technique for grade prediction that integrates both course- Four algorithms used in this paper:
wise and temporal influence effects. Courses and students are SVM Algorithm: Machine learning supports predicting and
portrayed in this approach as "knowledge" space. The classifying data, and we use a variety of machine learning
methods depending on the database. Support Vector
similarity of a student's latent representation in the
Machine, or SVM, is a straightforward model that can be used
"knowledge" space is modelled as a student's marks. The
to solve division and reversal tasks. It can solve both direct
course-by-course influence is taken into account as an extra and indirect problems and is useful for a variety of
component in predicting grades. The suggested method applications. SVM is a basic system: The method divides data
improves various standard approaches in inferring into categories by drawing a line or hyper plane.The radial
meaningful patterns between pairing of courses within basis function kernel, or RBF kernel, is a famous kernel
academic programmes, according to our results. function in machine learning that is utilised in a variety of
4. Divide students using Data mining algorithms: kernelized learning algorithms. It's especially popular in
support vector machine classification. Intuitively, the further
In this paper, we explore and compare algorithms with our data points are from the hyperplane, the more certain we
several data mining methods to differentiate students based are that they have been classified correctly. As a result, we
on their Moodle usage statistics and final course marks. We want our data points to be as far away from the hyperplane as
have developed a data mining application to make it easier for feasible while being on the correct side. Therefore, when new
educators to prepare and implement data mining techniques. test data is entered, the reaction phase is determined in which
We used real data from Cordoba University students in seven direction the hyperplane reaches.
Moodle studies. On original numerical data, we used Random Forest Algorithm: This is an ensemble algorithm,
discretization and rebalancing pre-processing approaches to which means it will develop an accurate classifier model by
combining different classifier methods. Internally, this
see if better classifier models could be obtained. Finally, we
technique will construct a train model for classification using
argue that a classifier model suitable for instructional
a decision tree algorithm.
purposes must be both accurate and understandable. Finally, Decision Tree Algorithm: This algorithm creates a training
we claim that, in order to be useful for decision making, a model by grouping comparable records together in the same
classifier model fit for instructional application must be both branch of the tree and continuing until all records are grouped
accurate and accessible to teachers. together in the complete tree. The classification train model
refers to the entire tree.
Gradient Boosting Classifiers: Gradient boosting classifiers
III. METHODOLOGY combines multiple weak learning models to make a powerful
predictive model. When making gradient boosting, decision
In this project author is describing concept to predict or trees are often used. Gradient enhancement models are
classify student performance based on their previous gaining popularity due to their ability to differentiate
academic performance. Using this paper, we will concentrate complex data sets, and have recently overcome many of
more on poor performance students by extracting grade Kaggle's data science challenges.
features from their past performance records. In this project Scikit-Learn, a Python machine learning package, offers a
we are using university dataset which contains record from A variety of gradient boosting classifier offerings, including
to W and we are extracting 4 features from this dataset to XGBoost.A single accurate train model will be built by
classify poor performing students. combining different techniques. Gradient Boosting gives
1. Taking those records from dataset which has features D better results in all of these algorithms.
and F and consider as failing student and we will assign
features values as 0 (Fgr) for such students.

2
 IV. PROPOSED SYSTEM

The majority of present approaches are aimed at identify

students who are at risk of failing to complete a course or
career and who could benefit from additional support. One
of the most important tasks in this process is to predict the
success of a student's grade. The downside to these methods
is that they do not work very well in predicting students who
are not doing well. We study the problem of forecasting a
student's performance at the conclusion of the semester
before he or she actually takes the course in this paper. To
focus on slower students who need these approaches.
Advantages of Proposed System:
It is easy to predict successful students, in order to have
better understanding.
Enhance the ability and success of their registered students.
Now click the 'Start Training Model' button to read the
V. RESULTS database and build a list of training objectives.

In the screenshot below, click the 'Upload Dataset Student

Grades Dataset' button to upload the database.

After uploading the data will go to the bottom of the screen

In the screenshot above we see the SVM accuracy is 69% and

we see the FSCORE value again. Now click on 'Run Random
Forest Algorithm' to create its own model.

3
In the screenshot above we can see total number of records in In the screenshot x-axis represents the name of the algorithm
dataset and then displaying algorithm chooses how many and the y-axis represents the accuracy.Now we can test new
records for training and testing purpose. Now click on student records on this train model to predict or classify new
‘Features Extraction’ button to extract features and assign as student performance. To check new student we need to
class label to the classifier algorithms. upload ‘text.txt’ test dataset from dataset folder and this
dataset contains below data.

In above test data we don’t have extracted feature values such

as 0,1,2,3 or 4 and this value will be predicted or classify by
In above screenshot we can see dataset contains total 14 this machine learning algorithms. Just we need to click on
numeric features and extracted features are 4. After feature ‘Classify Student Performance Reason’ button and upload
extraction click on ‘Run SVM Algorithm’ to build SVM train test dataset then will get below result.
model and to get classifier accuracy and FSCORE.

After uploading test data will get below classification result.

In above screen gradient boosting got 87% accuracy and

decision tree got high accuracy but its FSCORE is less
compare to gradient boosting. Now click on ‘Accuracy
Graph’ button to obtain accuracy graph.

In above screen based on grades values application has given

result as poor performance due to drop out or good
performance.

4
 VI. CONCLUSION

The purpose of this article is to accurately identify at-risk

students. These students may fail, drop out of class, or do
worse than usual. We used historical capture data to extract
features in order to test simple and advanced separation
methods based on large data sets.Based on AUC and F1 score
measures, the Gradient Boosting and Random Forest
classifiers perform the best. We have discovered several
interesting facts that help to explain student success.

VII. REFERENCES

1. B. Minaei-Bidgoli, D. A. Kashy, G. Kortemeyer, and

W. F. Punch:” Predicting student performance: an
application of data mining methods with an educational
web-based system”. In Frontiers in education, 2003.
FIE 2003 33rd annual, volume 1, pages T2A–13. IEEE,
2003.
2. Z. Iqbal, J. Qadir, A. N. Mian, and F. Kamiran: “A Case
Study on Machine learning based student grade
prediction”,Issue:2017.
3. J. E. Knowles. Of needles and haystacks: “Building an
accurate statewide dropout early warning system in
Wisconsin”. Journal of Educational Data Mining,
Issue:2015.
4. E. Osmanbegovic and M. Suljic:” Datamining approach
for predicting student performance”. Economic Review,
Issue: 2012.