Work

Construction of self-supervised learning models for educational scenarios based on data-efficient artificial intelligence

Year: 2026

Type: article

Abstract: This study proposes a Transformer-based self-supervised learning (SSL) framework for data-efficient educational intelligence under sparse and heterogeneous learner interaction conditions. The framework integrates masked educational interaction reconstruction, contrastive representation learning, temporal interaction-gap modeling, and Transformer-based contextual encoding. Experiments on EdNet used a learner-disjoint training/validation/test protocol to prevent cross-learner information leakage. In the full-label setting, the proposed SSL Transformer achieved an Accuracy of 0.5440, a Macro F1-score of 0.4772, and an AUC of 0.5225. Relative to the supervised Transformer baseline, these results represent improvements of approximately 3.5%, 4.1%, and 6.1%, respectively, although the proposed model did not maximize raw Accuracy across all baselines. With only 10% of the labeled training data, AUC remained 0.5203 versus 0.5225 under full supervision, while Accuracy and Macro F1-score were 0.4064 and 0.3978. Under the most restrictive cold-start condition of 1–3 observed interactions, the model achieved an Accuracy of 0.5063, a Macro F1-score of 0.4585, and an AUC of 0.5066. These findings support the value of self-supervised contextual representation learning under limited-label and sparse-history educational conditions.

Source: Scientific Reports

DOI: 10.1038/s41598-026-69479-2

Authors: Chou Chen, Yuehua Bai, Mani Amiri

Institutions: Wenzhou University of Technology, Wuhan Donghu University, Islamic Azad University of Tabriz

Language: English

Topic: Online Learning and Analytics

Subfield: Computer Science Applications

Field: Computer Science

Domain: Physical Sciences

SDG: Quality education

Open Access status: gold

Estimated APC paid: $2,690