Web & Social Media Analytics Previous Year Question Paper.pdf
Educational Data Mining Predicts Student Performance
1. Educational Data
Mining & Students’
Performance Prediction
By Amjad Abu Saa Information Technology Department Ajman University of Science and
Technology Ajman, United Arab Emirates
Paper Review by: Karishma Kuria
2. What is Educational Data Mining
EDM is emerging discipline of data mining, which deals with finding useful patterns and
discovering knowledgeable insights from vast variety of educational informational systems data.
The data is accumulated from various educational sources such as admissions systems, registration
systems, course management systems.
It focuses on developing new tools and techniques for discovering data insights.
EDM applies techniques from statistics, data mining algorithms, machine learning and AI to analyze
the data accumulated during teaching and learning.
3. Goals of EDM
To predict the performance of the students in future semesters and assist the educational institutes
and students to enhance their performance in future.
To help management improvise their teaching practices and financial activities to yield good results.
Identify the choices of student and to let them understand which areas they need to focus on to
improve their grades. This saves valuable time of both student and authorities.
Understanding all the factors which influence the performance of students and hence determining
the pedagogical support that can be provided by learning software.
4. Classification
Most famous, easy and widely used data mining technique.
It belongs to supervised learning technique, where target values are provided with the input data
In this technique we predict the class for each data element from the set of predefined class labels.
There are various classification techniques used in data mining for instance Neural Networks, Naïve
Bayes, K-Nearest Neighbour and Decision Trees.
In this paper 4 Decision trees and Naïve Bayes algorithm is used to design prediction models.
5. Data Mining Process
Dataset:
A survey conducted anonymously and without any bias on a group of from University of Science and
Technology (AUST), Ajman, United Arab Emirates.
Questionnaire consisted of several personal, academic and social questions which influences the
performance of student.
The data is further preprocessed, transformed to be appropriate for data mining techniques.
Grade Point Average (GPA) is used here to measure student performance with 4.0 being the
maximum.
Dataset consists of several attributes such as Student’s gender, Nationality, First Language, High
School percentage, Student Discount, Transportation etc.
8. Data Mining Implementation
& Results
What is Decision Tree?
A Decision tree is a supervised data
mining technique which is used to build
classification and predictive models. As
the name suggests it creates a top-down
model structured as tree from the
incoming dataset attributes, having a root
node and multiple incoming and outgoing
edges called Interior Nodes and Leaf
Nodes.
9. Data Mining Implementation & Results
In this paper following 4 Decision trees algorithm are used to design
prediction models.
C4.5 Decision Tree
ID3 Decision Tree
CART Decision Tree
CHAID Decision Tree
10. C4.5 Decision
Tree
Settings used:
Ø Splitting criterion = information gain
ratio
Ø Minimal size of split = 4
Ø Minimal leaf size = 1
Ø Minimal gain = 0.1
Ø Maximal depth = 20
Ø Confidence = 0.5
From the above matrix we can depict that out of 270 objects this algorithm predicted class of 95 objects and
provided an Accuracy value of 35.19%.
11. ID3 Decision
Tree
Settings used:
Ø Splitting criterion = information gain
ratio
Ø Minimal size of split = 4
Ø Minimal leaf size = 1
Ø Minimal gain = 0.1
From the above matrix we can depict that out of 270 objects this algorithm predicted class of 90 objects and
provided an Accuracy value of 33.33%.
12. Classification
and Regression
Tree
Settings used:
Ø Minimal leaf size = 1
Ø Number of folds used in minimal
cost-complexity pruning = 5Minimal
leaf size = 1
From the above matrix we can depict that out of 270 objects this algorithm predicted class of 108 objects and
provided an Accuracy value of 40%.
13. CHi-squared Automatic
Interaction DetectionTree
Settings used:
Ø Minimal size of split = 4
Ø Minimal leaf size = 2
Ø Minimal gain = 0.1
Ø Maximal depth = 20
Ø Confidence = 0.5
From the above matrix we can depict that out of 270 objects this algorithm predicted class of 92 objects and
provided an Accuracy value of 34.07%.
14. Analysis and Summary
Ø CART outperformed all the other
algorithms with an accuracy of 40%.
Which is significantly higher then the
expected accuracy.
Ø CHAID and C4.5 was next with 34.07%
and 35.19% respectively.
Ø The least accurate was ID3 with
33.33%.
15. Analysis and Summary
Ø Most of the algorithms scuffled in distinguishing similar class object.
Ø For instance in the class Pass of the CHAID matrix 25 out of 61 are
considered very good which comes in the upper 2 nearest class of grades
and 23 are considered in Pass category which comes in lower class of
grades.
Ø This refers that the discretization of class attributes was not independent
enough to capture the difference in other attributes.
Ø This confuses the model which deteriorates its accuracy and performance.
16. Naïve Bayes
Classification
Ø It is a probabilistic machine
learning model used in
classification tasks
Ø It assumes that there is no
dependencies between the
attributes in dataset.
Ø MOCS = Service: Interestingly, when the mother occupation status is on service, it appears that
students get higher grades.
Ø DISCOUNT: Students with higher grades tend to get discounts from the university more than low
grades students
19. Conclusion
As per the data mining pipeline the following steps were performed.
Data collection: Survey with Students
Data preprocessing and transformation
Data mining tasks, such as various Decision Trees and Naïve Bayes algorithm applied
Knowledgeable insights were drawn and performance was predicted
It can be inferred from the study in the paper that, student performance is not completely
dependent on their previous grades, there are other social and personal factor influencing their
performance