Source-linked AI summary
Data Mining: A Prediction for Performance Improvement of Engineering Students using Classification
Surjeet Kumar Yadav, Saurabh Pal
TL;DR
Educational data mining can use student records to predict examination outcomes and identify students needing support. This study constructs classification-based prediction models with decision-tree algorithms and reports accuracy for predicting pass, fail, and promotion outcomes, including 67.7778% accuracy for C4.5.
Problem
Educational databases contain information that can help predict examination outcomes and support low-achieving engineering students before examinations.
Method
The study constructs classification-based prediction models using decision-tree algorithms, including C4.5, ID3, and CART, with C4.5 supporting categorical and continuous attributes and missing values.
Results
67.7778% accuracy was achieved by C4.5, while ID3 and C4.5 achieved a 0.786 true positive rate for the FAIL class.
Takeaways & Limitations
The models can identify students likely to fail for counseling and produce prediction lists for incoming students needing special support.
Abstract
from arXiv · showhide
Now-a-days the amount of data stored in educational database increasing rapidly. These databases contain hidden information for improvement of students' performance. Educational data mining is used to study the data available in the educational field and bring out the hidden knowledge from it. Classification methods like decision trees, Bayesian network etc can be applied on the educational data for predicting the student's performance in examination. This prediction will help to identify the weak students and help them to score better marks. The C4.5, ID3 and CART decision tree algorithms are applied on engineering student's data to predict their performance in the final exam. The outcome of the decision tree predicted the number of students who are likely to pass, fail or promoted to next year. The results provide steps to improve the performance of the students who were predicted to fail or promoted. After the declaration of the results in the final examination the marks obtained by the students are fed into the system and the results were analyzed for the next session. The comparative analysis of the results states that the prediction has helped the weaker students to improve and brought out betterment in the result.
I. INTRODUCTION
Educational data mining applies knowledge-discovery methods to educational data, including classification for predicting student performance. This study frames predictive variables and classification models to assist low-achieving engineering students.
- Educational data mining discovers knowledge from data originating in educational environments using techniques including decision trees and Bayesian methods.
- Classification can support predictions about student performance and identify students likely to fail before examinations.
- The investigation aims to assist low academic achievers in engineering by generating predictive variables and identifying factors affecting learning behavior and performance.
- A prediction model is constructed using classification data-mining techniques based on identified predictive variables.
- The developed model is validated for engineering students studying in Indian universities or institutions.
II. DECISION TREE
Decision trees represent classification decisions as branching structures that recursively split data into outcomes. The section introduces ID3, C4.5, and CART as widely used decision-tree learning algorithms.
- A decision tree uses internal nodes for attribute tests, branches for distinct outcomes, and leaf nodes carrying class labels.
- Decision trees recursively split nodes from a root, with each branch representing a possible decision scenario and outcome.
- The three widely used decision-tree learning algorithms are ID3, C4.5, and CART.
A. ID3 (Iterative Dichotomiser 3)
The paper describes the construction and pruning characteristics of ID3, C4.5, and CART decision trees. Their attribute-selection measures and handling of continuous, missing, and unreliable data differ.
- ID3: ID3 selects splitting attributes using information gain and accepts categorical attributes, but noise can reduce its accuracy.
- ID3: ID3 builds trees by selecting the highest-information-gain root and recursively testing whether instances share a class.
- ID3: ID3 can handle continuous attributes through discretization or threshold-based splitting, but it does not support pruning.
- C4.5: C4.5 handles categorical, continuous, and missing attribute values, selects attributes using gain ratio, and reduces information-gain bias for attributes with many outcomes.
- C4.5: C4.5 uses pessimistic pruning to remove unnecessary branches and improve classification accuracy.
- CART: CART handles categorical, continuous, and missing values when building decision trees.
- CART: CART uses the Gini Index, produces binary splits, and applies cost-complexity pruning to remove unreliable branches.
III. BACKGROUND AND RELATED WORK
Prior educational data-mining studies use classification, regression, clustering, association rules, and decision trees to analyze or predict student performance. Reported applications span diverse institutions, variables, and outcomes.
- Educational data mining examines learning-process variables to improve understanding of education and student learning.
- Data-mining software can analyze data across dimensions, categorize it, and summarize relationships identified during mining.
- Bayesian classification studies linked student performance with academic, demographic, socioeconomic, family, and behavioral factors.
- Bayes classification was used to predict whether newcomer students would perform, using category, language, and background qualification variables.
- Regression analysis examined relationships between performance and attendance, study time, family income, and maternal characteristics.
- A study of cognition, personality, and demographic variables reported different academic-achievement patterns by gender and socioeconomic status.
- CHAID and CART classified successful and unsuccessful information-systems students from enrollment data with accuracies of 59.4 and 60.5 respectively.
- Earlier work used student learning data to warn at-risk students before final examinations and used decision trees to predict semester performance.
IV. DATA MINING PROCESS
Effective prediction of student performance requires considering many personal, social, psychological, environmental, and academic parameters.
- Prediction models should include personal, social, psychological, and environmental variables to predict student performance effectively.
A. Data Preparations
The study uses an initial dataset of 90 engineering students from VBS Purvanchal University for the 2010 session.
- 90 engineering-student records were initially collected from VBS Purvanchal University for the 2010 session.
B. Data selection and transformation
The study selects student background and performance attributes, defines their domain values, and uses first-year engineering results as a three-class response variable.
- B. Data selection and transformation: Student attributes include demographic, past-performance, address, contact, and derived variables selected from enrolment and database records.Most attributes represent students’ prior performance.
- B. Data selection and transformation: The paper defines domain values for categorical variables including category, high-school grade, senior-secondary grade, medium, and family size.High-school and senior-secondary grades use the O-to-F percentage mapping; medium is Hindi or English, and family size ranges from one to greater than three.
- B. Data selection and transformation: The dataset covers three engineering branches: Computer Science and Engineering, Information Technology, and Mechanical Engineering.
- B. Data selection and transformation: Admission type is represented by UPSEE admission or direct university admission.
- B. Data selection and transformation: The response variable is first-year engineering result, classified as Pass, Promoted, or Fail according to subject failures.Passing all papers yields Pass; specified combinations of theory and practical failures yield Promoted; otherwise the student is Fail.
C. Implementation of Mining Model
The implementation uses WEKA to apply ID3, C4.5, and CART decision-tree classifiers, with 10-fold cross-validation for model evaluation.
- C. Implementation of Mining Model: The engineering dataset is converted into an engg.arff file and loaded into the WEKA Explorer for classification.WEKA supports applying predictive algorithms, estimating model accuracy, and visualizing predictions or the model.
- C. Implementation of Mining Model: ID3, C4.5, and CART generate decision-tree predictive models evaluated using 10-fold cross-validation.Cross-validation is used because no separate evaluation dataset is available.
D. Results and Discussion
Three decision-tree algorithms generated predictive models and rules from the student dataset. Under 10-fold cross-validation, C4.5 achieved the highest reported accuracy, while Table IV reports the best class-wise accuracy models.
- ID3, C4.5, and CART generated decision-tree predictive models from the student dataset.
- Figures 1, 2, and 3 show the rules generated by ID3, C4.5, and CART, respectively.
- The generated rules classify students into outcomes including Pass, Fail, Promoted, and null cases based on attributes such as SSG, HSG, branch, and qualifications.
- C4.5 achieved the highest accuracy at 67.7778% under the reported 10-fold cross-validation comparison.
- Table III compares the execution time required by classifiers to build training models.
- Table IV identifies the three machine-learning algorithms producing the best class-wise accuracy.
V. CONCLUSIONS
The study evaluates decision-tree classifiers for predicting engineering students’ examination performance and reports useful identification of students likely to fail. It concludes that predictive models can support targeted interventions and improve student outcomes.
- Decision-tree classifiers were experimentally evaluated for predicting First Year engineering students’ examination performance.The study examines frequently used classifiers to identify a suitable model for student-performance prediction.
- 0.786 true positive rate for the FAIL class was achieved by both ID3 and C4.5 decision trees.This indicates successful identification of students likely to fail.
- Predictive models can produce short, accurate prediction lists for incoming students using records accumulated from previous years.The study also identifies students who may need special support.