A Data Engineering-Driven Machine Learning Framework for EFL Writing Performance Analysis

Explainable machine learning has been applied to textual predictors of writing quality, but rarely to the behavioral and motivational predictors, feedback modality and self-regulated learning (SRL), that dominate confirmatory EFL writing research. This study analyzes a quasi-experimental, ten-week EFL writing-feedback intervention (n=84), recording feedback condition (instructor-mediated, automated, or generative-AI), proficiency, and 35-item SRL strategy use alongside pre/post analytic writing scores across six outcomes. A pre-modeling reliability check found the SRL questionnaire did not reproduce its published factor structure (Cronbach's alpha = 0.279), so a single composite served as the motivational predictor, with centered feedback-by-SRL interaction terms. Five model classes were compared under nested cross-validation with per-trait tuning. Random forest achieved the best aggregate performance (MCRMSE = 2.240, mean R² = 0.293); a stacked ensemble performed worse than its base learners on every trait (mean R² = −0.344). A fixed-hyperparameter baseline outperformed the nested-CV-tuned pipeline (MCRMSE 2.197 vs. 2.258), showing tuning, not feature engineering, degraded performance here. SHAP attribution identified feedback condition, especially generative-AI feedback, as the dominant predictor for all six outcomes, with feedback-by-SRL interactions outweighing SRL's own main effect on three of six traits. A multi-task neural network outperformed an equivalent single-task ensemble on every trait using 5.75 times fewer parameters. These results show explainable, cross-validated machine learning can corroborate confirmatory EFL writing findings while surfacing measurement and methodological cautions.