Scholarlink

<p>The advancement in Artificial Intelligence (AI) is making room for the development of Smart Learning Environments (SLE). However, researchers have explored various ways Machine Learning (ML) can be leveraged to enhance the electronic learning experience. Nonetheless, over the past years, ML has been employed in the prediction of learners’ engagement. However, to ascertain the extent of breakthroughs experienced in this domain, this research collected 48 kinds of literature from seven scholarly databases, including IEEE and Springer, using suitable inclusion and exclusion criteria. The collected literature was reviewed with a focus on sources and nature of datasets, feature-sets utilization, feature extraction and selection techniques employed, and techniques employed in the development of models and their effectiveness. However, the findings from the review show that both shallow and deep learning techniques are employed in the development of learners’ engagement. Nonetheless, hybrid techniques are also leveraged for this purpose. Furthermore, it was discovered that CNN-based architectures are prevalently used. However, several public datasets are available online, and Datasets for Affective State in E-Environments (DAiSEE) is predominantly used. Nevertheless, learners’ emotional features are largely employed for this purpose. While researchers who employed deep learning techniques rarely employ feature selection, it was discovered that feature selection is commonly associated with shallow learning. Nonetheless, feature extraction is commonly used by researchers who utilize a deep learning approach. However, a detailed assessment of the predictive accuracies of the techniques shows that EfficientNet supplied the highest accuracy (99%). Nonetheless, a comparative analysis of different models developed using similar datasets shows that the quality and nature of the dataset utilized in the development of a model affect its performance. Furthermore, the research examined the impacts of engagement prediction models on learning outcomes. The review concluded by exposing limitations in the reviewed literature, which are proposed as directions for future research. The insight from this review will provide a clear directive for further investigation in this domain.</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

Shallow and deep learning approaches for predicting learners’ engagement in electronic learning systems: feature selection, extraction and engineering approaches

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • God’swill Theophilus,
  • Christopher Ifeanyi Eke,
  • Ovre Jeffrey Agushaka,
  • Timothy Moses

摘要

The advancement in Artificial Intelligence (AI) is making room for the development of Smart Learning Environments (SLE). However, researchers have explored various ways Machine Learning (ML) can be leveraged to enhance the electronic learning experience. Nonetheless, over the past years, ML has been employed in the prediction of learners’ engagement. However, to ascertain the extent of breakthroughs experienced in this domain, this research collected 48 kinds of literature from seven scholarly databases, including IEEE and Springer, using suitable inclusion and exclusion criteria. The collected literature was reviewed with a focus on sources and nature of datasets, feature-sets utilization, feature extraction and selection techniques employed, and techniques employed in the development of models and their effectiveness. However, the findings from the review show that both shallow and deep learning techniques are employed in the development of learners’ engagement. Nonetheless, hybrid techniques are also leveraged for this purpose. Furthermore, it was discovered that CNN-based architectures are prevalently used. However, several public datasets are available online, and Datasets for Affective State in E-Environments (DAiSEE) is predominantly used. Nevertheless, learners’ emotional features are largely employed for this purpose. While researchers who employed deep learning techniques rarely employ feature selection, it was discovered that feature selection is commonly associated with shallow learning. Nonetheless, feature extraction is commonly used by researchers who utilize a deep learning approach. However, a detailed assessment of the predictive accuracies of the techniques shows that EfficientNet supplied the highest accuracy (99%). Nonetheless, a comparative analysis of different models developed using similar datasets shows that the quality and nature of the dataset utilized in the development of a model affect its performance. Furthermore, the research examined the impacts of engagement prediction models on learning outcomes. The review concluded by exposing limitations in the reviewed literature, which are proposed as directions for future research. The insight from this review will provide a clear directive for further investigation in this domain.