Machine Learning for Educational Assessment: An Explainable AI Framework for Soft Skills Classification

The misalignment between university graduates' competencies and labor market demands represents a critical challenge in higher education. This study addresses this gap by integrating interpretable machine learning with online in-basket simulations to assess soft skills across educational levels. Using data from 611 students across four European universities, we applied eight machine learning algorithms with systematic hyperparameter optimization via GridSearchCV and 10-fold cross-validation. Gradient Boosting emerged as the optimal classifier, achieving 93.4% AUC and 80.98% accuracy in distinguishing Diploma, Bachelor's, and Master's students based on time management, planning, priority management, and field of study. Local Interpretable Model-Agnostic Explanations (LIME) revealed distinct classification mechanisms: Diploma students are characterized exclusively by field of study (impact = 0.998), Bachelor's students by time management (impact = 0.306), and Master's students by both dimensions (field = 0.554, time = 0.164). Perfect Diploma classification (100% accuracy) contrasts with Bachelor's-Master's overlap (27-22% mutual misclassification), indicating that vocational education produces field-specific profiles while university education cultivates convergent time management capabilities. These findings provide empirical evidence for targeted educational interventions, demonstrating that explainable AI combined with simulation-based assessment offers a scalable, transparent framework for competency evaluation aligned with workforce requirements.