Explainable TabNet for Academic Risk Prediction in Secondary Education: A Leave-One-Semester-Out Evaluation with Teacher-Centered Interpretability Framework
Year: 2026
Type: article
Abstract: Early identification of academically at-risk students is critical in resource-constrained Indonesian secondary schools, yet data-driven early warning systems remain scarce. This study develops a TabNet-based academic risk prediction framework integrated with four explainability methods, namely SHAP, LIME, attention masks, and counterfactual analysis, designed to produce teacher-comprehensible outputs from routine school records. Using Leave-One-Semester-Out temporal validation on longitudinal data from 1,248 students across six semesters at SMPN 1 Cibadak, Sukabumi, Indonesia, 21 features spanning academic performance, attendance, and extracurricular engagement were engineered into three composite risk indices. TabNet achieved an AUC-ROC of 0.9668, competitive with the best-performing XGBoost at 0.9887, with a non-statistically significant difference. At-risk recall reached 0.82, and SHAP validation confirmed attention mask reliability with a Spearman correlation of 0.7065. Extracurricular participation and total absences emerged as dominant predictors across all explainability methods. Counterfactual analysis of representative at-risk cases reveals meaningful variation in intervention pathways, distinguishing students amenable to targeted extracurricular re-engagement from those requiring more comprehensive multi-domain support. Practitioner validation by 18 homeroom teachers confirmed that the binary risk labels reflect educationally meaningful distinctions, with an overall percent agreement of 86.2% and Cohen’s kappa of 0.792, indicating substantial agreement between system-generated labels and independent professional judgement. The study demonstrates that TabNet, combined with multi-method explainability, effectively balances predictive accuracy with interpretability for non-technical educators. Deployed as an on-premises Indonesian-language dashboard, this framework offers a feasible early warning solution for Indonesian schools, pending formal teacher usability evaluation, cross-school validation, and equity analysis across student subgroups.
Language: English
Subfield: Computer Science Applications
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: green
OpenAlex ID: W7166539668
FWCI:
Cites:
Cited by:
Related to: