Source-linked AI summary
A Study on Feature Selection Techniques in Educational Data Mining
M. Ramaswami, R. Bhaskaran
TL;DR
Student-performance prediction depends on selecting relevant features, yet evidence on compact subsets and comparative filter-method quality is limited in this setting. The study evaluates six filter techniques on questionnaire data using NaiveBayes and ROC and F-measure criteria, finding best-performing methods and subset dimensionalities. Its results support using fewer features for predictive modeling while reducing training and classification costs.
Problem
The study addresses how to identify compact, relevant feature subsets and evaluate filter-selection quality for student-performance prediction.
Method
The study evaluates six filter feature-selection techniques on 1969 students’ questionnaire data, using NaiveBayes with ROC and F-measure evaluation.
Results
IG achieved the maximum ROC value of 0.729 with 7 features, while CH, IG, and SU each achieved an F1-Measure of 0.592 with 12 features.
Takeaways & Limitations
The study supports selecting a minimum number of features for student-performance models and reports reduced computational time and constructional cost in training and classification.
Abstract
from arXiv · showhide
Educational data mining (EDM) is a new growing research area and the essence of data mining concepts are used in the educational field for the purpose of extracting useful information on the behaviors of students in the learning process. In this EDM, feature selection is to be made for the generation of subset of candidate variables. As the feature selection influences the predictive accuracy of any performance model, it is essential to study elaborately the effectiveness of student performance model in connection with feature selection techniques. In this connection, the present study is devoted not only to investigate the most relevant subset features with minimum cardinality for achieving high predictive performance by adopting various filtered feature selection techniques in data mining but also to evaluate the goodness of subsets with different cardinalities and the quality of six filtered feature selection algorithms in terms of F-measure value and Receiver Operating Characteristics (ROC) value, generated by the NaiveBayes algorithm as base-line classifier method. The comparative study carried out by us on six filter feature section algorithms reveals the best method, as well as optimal dimensionality of the feature subset. Benchmarking of filter feature selection method is subsequently carried out by deploying different classifier models. The result of the present study effectively supports the well known fact of increase in the predictive accuracy with the existence of minimum number of features. The expected outcomes show a reduction in computational time and constructional cost in both training and classification phases of the student performance model.
1 INTRODUCTION
Educational data mining supports student-performance prediction, but model quality depends on selecting relevant variables and evaluating predictors with measures beyond accuracy. This study targets compact feature subsets for Indian higher-secondary student prediction.
- Educational data mining develops methods for discovering patterns in educational data and understanding students and learning settings.
- Prediction models can support identification of students with low or weak academic achievement and inform teachers, parents, and educational planners.
- Predictive performance depends strongly on selecting the most relevant variables from student datasets.
- ROC value and F-Measure are used to assess predictor goodness because classification accuracy depends heavily on class base rates.
- The study investigates minimum-cardinality feature subsets and evaluates six filter algorithms using F-measure and ROC values generated with NaïveBayes.
2 FEATURE SELECTION
Feature selection reduces the input space by identifying relevant variables, but optimal subset search becomes difficult as dimensionality grows. Filters offer a faster alternative to wrappers while prior student models often omitted feature-selection procedures.
- Feature selection chooses input-variable subsets by removing irrelevant or non-predictive features, improving learning efficiency, predictive accuracy, and result simplicity.
- Optimal feature-subset discovery is intractable and feature-selection problems have been proved NP-hard as dimensionality increases.
- Traditional feature selection generates candidate subsets, evaluates them, applies a stopping criterion, and validates the selected result.
- Wrappers use the learning algorithm to evaluate features, whereas filters use heuristics based on general data characteristics.
- Filters are comparatively faster than wrappers, while earlier student-performance studies reported classifier accuracy without performing feature selection.
3 DATA SOURCE AND PREDICTION OUTCOMES
The study models higher-secondary student outcomes from questionnaire-based demographic, family, socioeconomic, academic, and environmental information. It evaluates six filter methods to identify useful feature subsets for classification.
- Higher-secondary education bridges school education and higher-learning specializations, while academic performance reflects socioeconomic, psychological, and environmental factors.
- The dataset contains questionnaire responses from 1969 Tamil Nadu higher-secondary students, with 32 nominal predictive features and a pass/fail response variable.
- Feature selection combines attribute-subset evaluators with rank or Greedy search methods to identify the best feature sets.
- The study evaluates Correlation-based, Chi-Square, Gain-Ratio, Information-Gain, Relief, and Symmetrical Uncertainty attribute methods.
- F1-Measure values are used in the investigation to evaluate selected feature subsets.
4 RESULTS AND DISCUSSION
The study evaluates six filter feature-selection techniques across feature subsets of varying sizes using ROC and F1-Measure, with classifier benchmarking of selected subsets. IG reaches the highest reported ROC value at seven features, while CH, IG, and SU reach the highest F1-Measure at twelve features.
- Evaluation procedure: Feature subsets were evaluated iteratively from two features upward using ROC and F1-Measure because the ranking methods did not agree.The evaluation used multiple subsets generated from the ranked feature lists.
- ROC results: ROC plots show ROC values for feature subsets of different cardinalities, with maxima and corresponding subset sizes used to determine optimal subset dimensions.The number of attributes is plotted against ROC value, and peak values are summarized in Table 2.
- ROC results: ROC values are highest for CB and IG, with a peak ROC value of 0.729 for IG at seven features and for CB at nine features.The study identifies IG as having the optimal dimensionality for the higher-secondary-student dataset based on ROC.
- Classifier benchmarking: Benchmarking uses NaiveBayes, Voted Perceptron, OneR, and PART to evaluate five optimal subsets: CB-9, IG-7, CH-12, IG-12, and SU-12.The subsets are selected from the best-performing cardinalities identified by the ROC and F1-Measure analyses.
- F1-Measure results: F1-Measure plots show macro-averaged values against the number of attributes, emphasizing overall trends and separation among the top-performing curves.Peak F1-Measure values and their corresponding cardinalities are summarized in Table 3.
- F1-Measure results: F1-Measure values are highest for CH, IG, and SU, each reaching 0.592 with twelve features.These methods are consequently identified as having optimal dimensionality for the dataset under the F1-Measure criterion.
A NaiveBayes classifier can achieve relatively good per-
Four classifiers were compared across feature subsets, with IG-7 producing the strongest overall predictive performance. The study identifies IG-7 as the optimal dimensionality based on multiple evaluation measures.
- Four classifiers were evaluated against feature subsets generated by five feature-selection procedures.
- More than 89 percent predictive accuracy was achieved by Voted Perceptron and OneR.
- IG-7, the information-gain subset containing the top seven features, yielded the optimal feature-set dimensionality.The determination combined ROC value, F1-Measure, predictive accuracy, and RMSE.
5 CONCLUSION
The study compares six filter feature-selection algorithms to identify the best method and optimal feature-subset dimensionality. Its results support using fewer features for student-performance modeling while anticipating lower training and classification costs.
- Six filter feature-selection algorithms were compared to identify the best method and optimal feature-subset dimensionality.
- The study supports increased predictive accuracy when student-performance models use a minimum number of features.
- Reducing the feature set is expected to lower computational time and constructional cost during training and classification.