ICA-Bayesian Framework for Interpretable Educational Prediction: Disentangling Learning Behaviors with Causal Inference and Classroom Validation

Abstract

This research addresses the critical interpretability deficit hindering the adoption of AI in education by proposing a novel machine learning framework that integrates psychometrically-grounded feature disentanglement with pedagogically-aware behavioral causal inference. Theoretically grounded in educational psychology (e.g., Bloom's Taxonomy) and causal discovery principles, the framework innovatively decomposes entangled learning behaviors into semantically independent components and establishes temporally dynamic, pedagogically meaningful causal pathways. Through a rigorous, semester-long controlled experiment with 180 high school students, we demonstrate significant improvements over traditional black-box approaches: 92.3% feature importance alignment with expert judgment (vs. 61.7%), a 26.7% absolute increase in knowledge mastery (vs. 12.0%), and a 64% reduction in instructional planning time for teachers. Furthermore, the framework demonstrated strong equity implications, reducing the achievement gap between top and bottom quartiles by 18%, with particularly pronounced benefits for low-performing students and reflective learners. This work provides educators with actionable, evidence-based decision support tools while addressing systematic research gaps in interpretable educational AI.