Scholarlink

Accurately predicting student academic performance remains a persistent challenge in educational data mining, particularly when dealing with limited, imbalanced, and noisy datasets. The present study proposes hybrid models in which denoising autoencoders (DAEs) are employed for feature extraction, followed by a range of deep learning classifiers—namely recurrent neural networks (RNN), long short-term memory networks (LSTM), artificial neural networks (ANN), deep neural networks (DNN), gated recurrent units (GRU), and convolutional neural networks (CNN)—to improve classification accuracy across multiple performance levels. In addition, a Stacking ensemble strategy is employed to integrate the strengths of individual models and enhance overall robustness. The experimental evaluation was conducted on a real-world educational dataset, transformed into a multi-class classification task reflecting four academic performance categories. The results demonstrate that the Stacking model achieved the highest accuracy (98.45%), followed by the DAE_RNN (Denoising Autoencoder with Recurrent Neural Network) and DAE_GRU (Denoising Autoencoder with Gated Recurrent Unit) models, which also exhibited strong class-wise F1-scores. Recurrent models consistently outperformed feedforward and convolutional counterparts, highlighting the importance of temporal pattern modeling in academic data. The proposed approach demonstrates the effectiveness of combining denoising-based representation learning with deep neural architectures to address classification challenges in the educational domain, offering a scalable and robust solution for learning analytics applications.

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

Stacked Deep Learning Models Leveraging Denoising Autoencoder-Based Features for Student Performance Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Amani Khalifa,
  • Fatma BenSaid,
  • Yessine Hadj Kacem

摘要

Accurately predicting student academic performance remains a persistent challenge in educational data mining, particularly when dealing with limited, imbalanced, and noisy datasets. The present study proposes hybrid models in which denoising autoencoders (DAEs) are employed for feature extraction, followed by a range of deep learning classifiers—namely recurrent neural networks (RNN), long short-term memory networks (LSTM), artificial neural networks (ANN), deep neural networks (DNN), gated recurrent units (GRU), and convolutional neural networks (CNN)—to improve classification accuracy across multiple performance levels. In addition, a Stacking ensemble strategy is employed to integrate the strengths of individual models and enhance overall robustness. The experimental evaluation was conducted on a real-world educational dataset, transformed into a multi-class classification task reflecting four academic performance categories. The results demonstrate that the Stacking model achieved the highest accuracy (98.45%), followed by the DAE_RNN (Denoising Autoencoder with Recurrent Neural Network) and DAE_GRU (Denoising Autoencoder with Gated Recurrent Unit) models, which also exhibited strong class-wise F1-scores. Recurrent models consistently outperformed feedforward and convolutional counterparts, highlighting the importance of temporal pattern modeling in academic data. The proposed approach demonstrates the effectiveness of combining denoising-based representation learning with deep neural architectures to address classification challenges in the educational domain, offering a scalable and robust solution for learning analytics applications.