Scholarlink

<p>The ability to predict and intervene on at-risk students early within programming courses where knowledge can become problematic quickly is paramount. Educational datasets tend to be small, imbalanced, and vulnerable to shifts among cohorts. To address the limitations of classical approaches in such conditions, this study aims to advance at-risk student predictions in programming courses through the design and implementation of a novel hybrid deep-learning pipeline that combines controlled-data augmentation, deep-learning embedding generation, and classical-machine learning classification. We experiment with a real university dataset containing 177 processed records of students from three C-language programming classes: Algorithm and Programming 1, Algorithm and Programming 2, and Data Structures. After pre-processing of academic and behavioral features, SMOTE is performed on the training dataset alone to reduce imbalance, while the test dataset stays untouched, retaining its raw student records. Two types of hybrid pipelines are created using a CNN-based and an ANN-based approach to the extraction of embeddings that are further used by several classifiers. It is shown experimentally that the hybrid models significantly outperform the baseline approaches in their accuracy and performance metrics of recall and F1-score, which are essential for recognizing at-risk students. In particular, hybrid models reach F1-scores of 0.93 for Target 1, 0.96 for Target 2, and 0.97 for Target 3. The confusion matrix analysis proves the models’ capability of accurately identifying the majority of at-risk students, while the paired t-tests performed after random splits confirm statistically significant improvement over the respective baselines in all three targets. We conclude that our research may be promising as a basis for model deployment and continued monitoring over subsequent academic years in order to detect any drift.</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 hybrid deep–machine learning framework with data augmentation for student outcome prediction in programming courses

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月23日

查看全文
  • Zakaria Soufiane Hafdi,
  • Said El Kafhali,
  • Oumaima Ghandour

摘要

The ability to predict and intervene on at-risk students early within programming courses where knowledge can become problematic quickly is paramount. Educational datasets tend to be small, imbalanced, and vulnerable to shifts among cohorts. To address the limitations of classical approaches in such conditions, this study aims to advance at-risk student predictions in programming courses through the design and implementation of a novel hybrid deep-learning pipeline that combines controlled-data augmentation, deep-learning embedding generation, and classical-machine learning classification. We experiment with a real university dataset containing 177 processed records of students from three C-language programming classes: Algorithm and Programming 1, Algorithm and Programming 2, and Data Structures. After pre-processing of academic and behavioral features, SMOTE is performed on the training dataset alone to reduce imbalance, while the test dataset stays untouched, retaining its raw student records. Two types of hybrid pipelines are created using a CNN-based and an ANN-based approach to the extraction of embeddings that are further used by several classifiers. It is shown experimentally that the hybrid models significantly outperform the baseline approaches in their accuracy and performance metrics of recall and F1-score, which are essential for recognizing at-risk students. In particular, hybrid models reach F1-scores of 0.93 for Target 1, 0.96 for Target 2, and 0.97 for Target 3. The confusion matrix analysis proves the models’ capability of accurately identifying the majority of at-risk students, while the paired t-tests performed after random splits confirm statistically significant improvement over the respective baselines in all three targets. We conclude that our research may be promising as a basis for model deployment and continued monitoring over subsequent academic years in order to detect any drift.