Design of an iterative model for educational video classification using graph-based self-training methods

Abstract

The critical need for advanced approaches in classifying educational videos has been extensively researched, as the domain has faced significant challenges due to the limited availability of labeled data, noisy annotations, and the inherent diversity of video content. Traditional approaches are likely to fall short in managing such complexity, resulting in suboptimal classification performance. A new integration of graph-based semi-supervised learning (GSSL), self-training with consistency regularization, adversarial learning, transfer learning from pretrained models, and a weakly supervised learning framework is proposed in this paper. All these approaches help improve performance in our proposed framework across several metrics, including increased precision, accuracy, recall, and the area under the curve (AUC), which subsequently reduces delays and increases specificity with respect to the EDUVSUM and HowTo100M datasets. The uniqueness of combining these techniques will also enhance the classification accuracy and competence of such models, resulting in a robust and generalizable classifier across various domains of educational content. This paper presents a significant contribution to the field of educational video classification by providing a comprehensive solution to the multifaceted challenges of the task.