Scholarlink

This study predicts student performance in higher education using a dataset of 226 engineering students, expanded to 50,000 samples using the synthetic minority oversampling technique (SMOTE). We propose a hybrid framework combining Random Forest (RF), Feedforward Neural Network with Attention (FNN+Attention), and XGBoost with Local Interpretable Model-agnostic Explanations (LIME) to classify performance into five categories: bad (<2.8), poor (2.8–3.0), average (3.0–3.2), good (3.2–3.4), and excellent (>3.4). RF achieved 88% precision (macro F1 = 0.87), FNN+Attention 90% (F1 = 0.89), and XGBoost 89% (F1 = 0.88). The key predictors, stress (Q9), sports participation (Q10), and social media usage (Q85), were identified using SHAP, attention weights, and LIME, allowing tailored recommendations (e.g., stress counseling). The novelty lies in scaling small educational datasets with SMOTE, integrating multidimensional predictors (academic, psychological, lifestyle), and combining machine learning, deep learning, and explainable AI for actionable insights in higher education.

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 Hybrid XAI Approach to Predict and Interpret Student Outcomes in Education

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pravesh Kumar Bansal,
  • Mushtaq Ahmed

摘要

This study predicts student performance in higher education using a dataset of 226 engineering students, expanded to 50,000 samples using the synthetic minority oversampling technique (SMOTE). We propose a hybrid framework combining Random Forest (RF), Feedforward Neural Network with Attention (FNN+Attention), and XGBoost with Local Interpretable Model-agnostic Explanations (LIME) to classify performance into five categories: bad (<2.8), poor (2.8–3.0), average (3.0–3.2), good (3.2–3.4), and excellent (>3.4). RF achieved 88% precision (macro F1 = 0.87), FNN+Attention 90% (F1 = 0.89), and XGBoost 89% (F1 = 0.88). The key predictors, stress (Q9), sports participation (Q10), and social media usage (Q85), were identified using SHAP, attention weights, and LIME, allowing tailored recommendations (e.g., stress counseling). The novelty lies in scaling small educational datasets with SMOTE, integrating multidimensional predictors (academic, psychological, lifestyle), and combining machine learning, deep learning, and explainable AI for actionable insights in higher education.