Model interpretability on private-safe oriented student dropout prediction.

Area of Science:

• Artificial Intelligence
• Education Management
• Data Science

Background:

• Student dropout poses significant societal challenges, impacting employability and sustainable development.
• Predicting dropout risks is crucial for educational interventions but faces privacy and interpretability issues with current AI models.
• Existing machine learning models often require real student data, raising privacy concerns and lacking transparency.

Purpose of the Study:

• To develop a privacy-preserving and interpretable AI framework for predicting student dropout.
• To address the limitations of traditional data synthesis and opaque machine learning models in educational contexts.
• To enhance the practical application of AI in sustainable education management.

Main Methods:

• Introduced a modified Preprocessed Kernel Inducing Points data distillation technique (PP-KIPDD) for synthesizing student data.
• Utilized PP-KIPDD to create privacy-protected training datasets, mitigating information leakage risks.
• Employed SHAP (SHapley Additive exPlanations) values to enhance model interpretability and feature significance.

Main Results:

• The PP-KIPDD technique demonstrated superior performance and efficiency compared to Conditional Generative Adversarial Networks for data synthesis.
• The approach successfully prevented student privacy information leakage.
• Enhanced model interpretability provided clear insights into features driving dropout predictions.

Conclusions:

• The novel AI framework offers a privacy-preserving and credible solution for student dropout prediction in sustainable education.
• This approach enhances the feasibility and reasonableness of AI applications in educational management.
• The study presents a new end-to-end framework for AI in sustainable education, prioritizing decision-maker needs for practical implementation.