Machine-Learning Classification of Archived Employability-Readiness Ratings: Performance, Calibration, and Internal Validation

Tonglu Men (Department of Education and Society, Rajamangala University of Technology Krungthep, Bangkok, Thailand)
Premsuree Chaumthong (Department of Education and Society, Rajamangala University of Technology Krungthep, Bangkok, Thailand)

International Journal of Environment, Engineering, and Education
Vol. 8 No. 3 (2026)

This study examined the explanatory structure and predictive reproducibility of an end-of-program administrative score used to assess graduate preparedness in vocational education. A cross-sectional secondary analysis was conducted using 450 de-identified learner records collected from four anonymized polytechnic-style institutions during the 2023–2024 academic year. The outcome was a 0–100 composite score analyzed continuously and as three prespecified categories: low (<60), moderate (60–79), and high (≥80). Candidate predictors comprised practical skills, digital competence, soft skills, internship intensity, project performance, and an industry–education integration index treated as an institution-linked contextual proxy. Pearson correlations, heteroskedasticity-consistent ordinary least-squares regression, criterion-overlap sensitivity analysis, and leakage-controlled model development were applied. Six algorithms were compared against a majority-class baseline using a stratified 70:30 development-test split. Bootstrap intervals, ordinal-error measures, and probability-calibration indices were used to quantify uncertainty and performance. The full regression model explained 76% of outcome variance (adjusted R² = 0.75), whereas the reduced model retained an adjusted R² of 0.58 after removal of predictors susceptible to shared rubric content. In the untouched test partition (n = 135), the back-propagation neural network achieved an accuracy of 0.904 and a macro-F1 of 0.903; all 13 errors occurred between adjacent categories. These findings indicate a coherent within-system scoring structure and strong reproducibility across analytical approaches. However, they do not establish causal effects, transportability across institutions, fairness, operational utility, or prediction of subsequent labor-market outcomes. The model should therefore be restricted to low-stakes auditing, data-quality review, and identification of borderline records pending prospective multi-institutional evaluation.