Machine learning driven multidimensional evaluation system for teaching quality of vocational education teachers

Abstract

Evaluating teaching quality in vocational education (VE) requires a multidimensional approach to align with industry-focused learning outcomes. Traditional single-metric scoring or subjective rubrics often overlook critical aspects such as pedagogy, safety, and technology integration. To address these limitations, we propose a Gradient Jellyfish Search-enriched LightGBM (GJS-ELGBM) framework that provides a comprehensive assessment of vocational instructors’ performance. The model analyzes seven domains—pedagogical skill, learner engagement, learning outcomes, innovation, professionalism, safety, and contextual factors—and assigns a four-tier quality label. Categorical features are one-hot encoded, and numeric features are normalized using Z-scores. t-SNE is applied for visual cluster validation without affecting the predictive pipeline. GJS-ELGBM combines global meta-heuristic exploration with local gradient refinement to efficiently optimize LightGBM hyperparameters and improve predictive accuracy. Statistical analysis, including one-way ANOVA with post-hoc Tukey HSD tests, confidence intervals, and effect sizes, was performed to validate the robustness and significance of the results. Experimental results demonstrate that the proposed GJS-ELGBM model achieves a high accuracy of 99.5% in this experimental setup, with SHAP analysis identifying classroom observation scores, technology integration, and pass rates as the most influential factors. The proposed framework offers an accurate, transparent, and computationally efficient tool for data-driven teacher evaluation, facilitating meaningful quality improvements across vocational programs.