Explainable AI in education: integrating educational domain knowledge into the deep learning model for improved...

Area of Science:

• Educational Technology
• Artificial Intelligence in Education
• Machine Learning for Learning Analytics

Background:

• Deep learning models like Artificial Neural Networks (ANNs) are prevalent for predicting student performance.
• The
• black-box
• nature of ANNs often yields unreliable insights inconsistent with educational domain knowledge.
• This lack of interpretability hinders model trustworthiness and performance optimization.

Purpose of the Study:

• To develop an interpretable ANN for student performance prediction.
• To address the inconsistency between ANN-learned relationships and established educational domain knowledge.
• To improve the accuracy and reliability of student performance prediction models.

Main Methods:

• Utilized Shapley Additive Explanations (SHAP) to analyze an ANN trained on Portuguese high school students' mathematics performance data.
• Developed the Student Performance Prediction Explanation (SPPE) algorithm to optimize ANN by incorporating educational domain knowledge.
• Conducted global and local interpretability analyses to track feature contribution changes.

Main Results:

• Identified key features influencing student mathematics performance.
• The original ANN model's learned correlations contradicted educational domain knowledge.
• The SPPE-optimized ANN demonstrated a 26.9% improvement in prediction accuracy compared to the original model.
• The optimized ANN outperformed traditional machine learning algorithms.
• The SPPE strategy showed robustness across different ANN architectures.

Conclusions:

• Integrating educational domain knowledge significantly enhances the accuracy and interpretability of ANN models for student performance prediction.
• The proposed SPPE algorithm offers a practical and generalizable approach for developing trustworthy AI in education.
• Findings provide actionable insights for improving educational applications and developing interpretable neural network frameworks.