Beyond Black Boxes: Interpretable AI for Academic Performance Prediction Using Advanced Tabular Neural Networks with Augmented Attention Mechanism

Lack of explainability in deep learning models remains a major barrier to their adoption in educational settings, where transparent decision-making is essential. This study presents ATAAP (Augmented Tabular Attention for Academic Performance Prediction), a novel interpretable neural network architecture that integrates an augmented attention mechanism within a TabTransformer framework to achieve high predictive performance while maintaining transparency. Unlike conventional explainability approaches that rely on post-hoc analysis, ATAAP incorporates explainability directly into the prediction process through augmented multi-head attention layers that assign importance weights to input features based on their contribution to model outputs. Two key innovations support this framework: a domain-aware initialization strategy that embeds educational feature-cluster knowledge into attention heads, and an attention-fidelity loss function that aligns learned attention weights with SHAP-derived feature importance scores obtained from a calibration dataset. The proposed model was evaluated using the UCI Student Performance dataset, the Open University Learning Analytics Dataset (OULAD), and a proprietary dataset from two Nigerian tertiary institutions (NAPP-23). ATAAP achieved classification accuracies of 94.7%, 88.3%, and 91.6%, respectively, with an F1-score of 0.943 and an AUC-ROC of 0.971 on the UCI benchmark. Furthermore, the model generated explanations that demonstrated greater fidelity to ground-truth feature attributions than competing interpretable and post-hoc methods. Fairness assessments across gender, socioeconomic status, and first-generation status revealed no systematic bias against protected groups. These findings demonstrate that ante-hoc attention-based explainability can deliver both robust predictive accuracy and faithful, instance-level explanations for high-stakes educational decision-making