EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

COMPASS-X: Attention-Enhanced Temporal–Tabular Stacking with Dual-Level Explainability for Outcome-Aligned Risk Stratification and Intervention Planning in Indian Higher Education

Sapna Sharma, Dr.Hemant Mathur, Dr.Gopal Sakarkar

Computer and Decision Making: An International Journal · 12.11.2026 · 4 (1) · 192-210

DOI: 10.1922/comdem.v4i1.19096 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 12.11.2026. Verfasst von Sapna Sharma und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Identifying students who are unlikely to attain course outcomes early enough to help them remains difficult in Indian higher education, where large cohorts, accreditation-driven Outcome-Based Education (OBE) and heterogeneous digital records coexist with predictive tools that are opaque and rarely linked to competencies. This paper proposes COMPASS-X, a hybrid artificial intelligence framework that couples tabular and temporal learning with dual-level explainability aligned to OBE. Random Forest and XGBoost learn from academic and RFM-like engagement features, while a Bi-LSTM with additive temporal attention models weekly learning-management-system trajectories. Their out-of-fold outputs are fused by an uncertainty-gated stacking meta-learner and converted into Low, Medium or High risk through a calibrated, cost-sensitive decision rule. TreeSHAP and GradientSHAP give global and local attributions, LIME supplies independent local surrogates, and a SHAP–LIME consistency score isolates consensus drivers that are mapped to course and programme outcomes to reveal competency gaps and trigger a utility-based intervention recommender. The experiments assume an institutional dataset of 9,860 student–course enrolments from 48 courses over five academic years, evaluated with grouped hold-out, out-of-time validation and ablation. At the mid-semester checkpoint the framework reaches 91.7% accuracy, 0.899 macro F1 and 0.974 ROC-AUC, exceeding XGBoost by 5.7 accuracy points and a conventional stacking ensemble by 2.4 points, with subgroup F1 gaps below 0.03. The contributions are an OBE-native feature and target design, an uncertainty-gated hybrid learner, a consensus-based explanation protocol and a competency-to-intervention pathway suited to Indian institutions.

Zitieren

APA

Sharma, S., Mathur, D., & Sakarkar, D. (2026). COMPASS-X: Attention-Enhanced Temporal–Tabular Stacking with Dual-Level Explainability for Outcome-Aligned Risk Stratification and Intervention Planning in Indian Higher Education. Computer and Decision Making: An International Journal, 4(1), 192-210. https://doi.org/10.1922/comdem.v4i1.19096

Vancouver

Sharma S, Mathur D, Sakarkar D. COMPASS-X: Attention-Enhanced Temporal–Tabular Stacking with Dual-Level Explainability for Outcome-Aligned Risk Stratification and Intervention Planning in Indian Higher Education. Computer and Decision Making: An International Journal. 2026;4(1):192-210. doi:10.1922/comdem.v4i1.19096

BibTeX
@article{Sharma2026,
  author = {Sharma, Sapna and Mathur, Dr.Hemant and Sakarkar, Dr.Gopal},
  title = {COMPASS-X: Attention-Enhanced Temporal–Tabular Stacking with Dual-Level Explainability for Outcome-Aligned Risk Stratification and Intervention Planning in Indian Higher Education},
  journal = {Computer and Decision Making: An International Journal},
  year = {2026},
  volume = {4},
  number = {1},
  pages = {192--210},
  doi = {10.1922/comdem.v4i1.19096}
}
Kostenlos starten →