Scholarlink

<p>Early prediction of students’ performance is essential and vital for institutions to improve academic achievement. The Kaggle dataset comprises 5,000 students’ demographic, academic, behavioral, socioeconomic, and well-being attributes used in framework development. The SMOTE, ROS, and SMOTE-ENN resampling methods used in handling the data imbalance. In this framework, constructed and tested performance applying six machine learning techniques (i.e., DT, RF, SVM, GB, FFNN, and ANN) through implementation. The model performance comparison using the metrics precision, recall, and F1-score of the Fail class, and also used stratified 5-fold Cross-Validation and the random hold-out method for model validation. The result shows that data resampling substantially improves the minority-class classification accuracy. A unified non-parametric evaluation framework combining Friedman ranking, Wilcoxon signed-rank testing, and effect-size analysis statistically ensures the best model. The rank-based tests, distribution-aware analysis, and practical effect sizes reveal that the RF with ROS has the highest Fail-class F1-scores, higher recall, and accuracy. The RF-based model is the most reliable model for early identification of at-risk of dropout. The proposed model can be used low performing students identification in educational institutions for timely intervention, motivation and additional support. This model also supports data-driven decision-making in support of students’ performance enhancement.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Novel Framework for Early Prediction of Students’ Performance

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Nisha Madaan,
  • Sanjay Kumar Sharma

摘要

Early prediction of students’ performance is essential and vital for institutions to improve academic achievement. The Kaggle dataset comprises 5,000 students’ demographic, academic, behavioral, socioeconomic, and well-being attributes used in framework development. The SMOTE, ROS, and SMOTE-ENN resampling methods used in handling the data imbalance. In this framework, constructed and tested performance applying six machine learning techniques (i.e., DT, RF, SVM, GB, FFNN, and ANN) through implementation. The model performance comparison using the metrics precision, recall, and F1-score of the Fail class, and also used stratified 5-fold Cross-Validation and the random hold-out method for model validation. The result shows that data resampling substantially improves the minority-class classification accuracy. A unified non-parametric evaluation framework combining Friedman ranking, Wilcoxon signed-rank testing, and effect-size analysis statistically ensures the best model. The rank-based tests, distribution-aware analysis, and practical effect sizes reveal that the RF with ROS has the highest Fail-class F1-scores, higher recall, and accuracy. The RF-based model is the most reliable model for early identification of at-risk of dropout. The proposed model can be used low performing students identification in educational institutions for timely intervention, motivation and additional support. This model also supports data-driven decision-making in support of students’ performance enhancement.