Beyond Accuracy: An Explainable AI-Driven Student Academic Performance Prediction System

This article introduces a student academic performance prediction study combining deep learning (LSTM, Bi-LSTM, diffusion models) and Explainable AI (SHAP, LIME, DICE, etc.), while integrating the fairness assessment tool fairsynedu, aiming to build an educational AI system that balances accuracy and explainability. The project source code and related materials are available on GitHub (link: https://github.com/Istiaq-Alam/Student-Performance-Prediction-with-Explainable-AI).

Traditional student academic prediction models often focus on accuracy but ignore decision transparency and explainability, making it impossible for educators to take effective interventions based on prediction results. The core question of this study: How to build an AI system that can accurately predict student performance while providing explainable and actionable insights?

Deep Learning Models: Uses LSTM (captures long-term dependencies in time series), Bi-LSTM (bidirectional context), diffusion models (data augmentation and synthetic samples); Explainable AI Methods: SHAP (feature contribution), LIME (single prediction explanation), DICE (diverse counterfactuals), CEM (minimum perturbation counterfactuals); Fairness Assessment: Uses the fairsynedu tool to detect biases in educational data and avoid systemic discrimination.

The study uses the classic Student Performance Data Set (Cortez and Silva, 2008), which includes math and Portuguese language grades data of Portuguese middle school students, covering dimensions: demographic characteristics (age, gender), socioeconomic data (parental education/occupation), behavioral characteristics (study time, extracurricular activities), and historical performance (past grades, absences).

The project develops an interactive dashboard based on Flask API, with features including: 1. Prediction function (input student features to get performance predictions); 2. Explainability display (visualization of SHAP values, LIME explanations); 3. Human-machine collaborative feedback (educators can provide feedback on prediction results to continuously optimize the model).

Academic Contributions: 1. Introduce diffusion models into the field of educational prediction; 2. Systematically compare the performance of multiple XAI methods in educational scenarios; 3. Incorporate fairness assessment into the model development process. Practical Value: 1.Early warning of students at academic risk;2.Provide personalized learning recommendations;3.Provide decision support for educational managers.

Current Limitations: Geographical limitation of the dataset (Portugal), high computational cost of counterfactual explanations, controversial selection of fairness metrics. Future Directions: Multimodal data fusion (learning logs/homework patterns), real-time prediction systems, cross-cultural validation, evolution from correlation to causal inference.