Explainable Deep Representation Learning for Digital Academic Resilience Classification in Higher Education
Year: 2026
Type: conference-paper
Abstract: Digital learning environments require students to demonstrate autonomy, self-regulation, emotional control, and persistence when facing academic and technological challenges. However, conventional psychometric approaches may not fully capture the nonlinear relationships, response variability, and measurement noise embedded in self-reported resilience data. This study proposes an interpretable deep representation learning framework for classifying digital academic resilience in higher education. The framework integrates Structural Equation Modelling (SEM) to validate the Online Academic Resilience Scale (OARS), a Denoising Autoencoder (DAE) to extract compact and noise-resistant latent features, an Artificial Neural Network (ANN) to classify resilience levels, and SHAP-based explainability to interpret the learned representations. The analysis used responses from 300 undergraduate students enrolled in distance-learning programmes. The DAE compressed the 20 OARS items into five latent features, which were then evaluated against raw questionnaire inputs across several classification models. The ANN-DAE configuration achieved the strongest performance, with an accuracy of 0.900, an F1-score of 0.900, and an AUC of 0.928, outperforming models trained on raw data. SHAP analysis further showed that latent dimensions associated with Digital Self-Regulation, Emotional Control in Digital Learning, and Digital Academic Confidence contributed most strongly to the classification results. These findings indicate that digital academic resilience is better represented as a multidimensional latent construct than as a simple aggregation of questionnaire scores. The proposed framework offers a transparent basis for early identification and targeted support of students at risk in digital learning environments.
Authors: Nyoman Ayu Gita Gayatri, Dian Kusumajati
Institution: Binus University
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: gold
OpenAlex ID: W7204529970