Analysis of Student Performance Prediction Using the Bi-LSTM Technique Focused on the Indian Higher Education Context Using a Synthetic Dataset

Abstract

Student performance prediction is a critical task for improving educational outcomes and enabling data-driven decision-making in higher education. This study proposes a context-specific predictive framework for Indian higher education that integrates AISHE institutional indicators with student-level behavioral data through multi-level data fusion, incorporates SHAP-based interpretability to support policy and intervention decisions, and employs a Bi-LSTM + Nadam architecture to achieve strong predictive performance. The framework integrates multi-level heterogeneous data, combining institutional indicators from AISHE with student-level academic, behavioral, and demographic features. To address the lack of unified large-scale datasets, a synthetic data generation approach is employed to simulate multi-institutional environments with temporal student records, supported by a mapping mechanism aligning student and institutional attributes. Experimental results demonstrate that the proposed model achieves 93.1% accuracy, significantly outperforming traditional machine learning approaches by approximately 13%. The model maintains precision of 93.0%, recall of 93.1%, and F1-scores of 93% across all performance categories, confirming robust and balanced classification. Convergence analysis shows that Nadam enables faster and more stable training, while the Bi-LSTM architecture improves performance by 1.9% over unidirectional LSTM by capturing bidirectional temporal dependencies. Feature importance analysis reveals that academic and behavioral factors dominate prediction, with GPA, attendance, and assignment submission contributing approximately 42.7% of the total normalized SHAP importance, while demographic factors have minimal influence, ensuring reduced bias. ROC-AUC analysis confirms excellent discrimination across all categories, with a macro-average AUC of 0.993 (range: 0.989–0.997), fully consistent with the reported accuracy and recall metrics. The proposed framework is suitable for early warning systems, targeted interventions, and academic decision support, making it highly applicable in real-world educational settings.