POA-MLSP: a multi-dimensional learning analytics framework for predicting CET4 writing performance based on a production-oriented approach and student engagement patterns

Abstract

Contemporary College English Test Band 4 (CET-4) writing instruction faces significant challenges in accurately predicting student performance and providing timely pedagogical interventions. This study develops and validates the Production-Oriented Approach Multi-Dimensional Learning Analytics Framework for Student Performance (POA-MLSP) for predicting CET-4 writing performance across five dimensions through systematic integration of Production-Oriented Approach (POA) theory and Self-Determination Theory (SDT)-based engagement modeling. The framework implements a four-layer architecture incorporating Feature Adaptive Selection Mechanism and SDT-Based Engagement Dynamic Modeling algorithms. Validation involves 124 students during a 16-week semester, collecting multi-source data including Jacobs' five-dimensional assessments, Utrecht Work Engagement Scale-Student (UWES-S) engagement measurements, classroom observations, and digital platform interactions across experimental and control groups. POA-MLSP achieves R² = 0.75 overall prediction accuracy, outperforming linear regression (R² = 0.58), random forest (R² = 0.66), and support vector machines (R² = 0.63) by 17-29%. Content prediction reaches highest accuracy (R² = 0.78), while the framework identifies five distinct engagement profiles and achieves 78.4% ± 2.1% early warning accuracy with 79.8% ± 2.9% teacher satisfaction. Educational theory-guided algorithms significantly enhance prediction performance while maintaining pedagogical interpretability, enabling proactive intervention through early warning systems with minimal implementation burden for authentic educational applications.