Аннотация
The study examines the use of a machine learning framework that predicts the preparation level of prospective mathematics teachers through the integration of academic, behavioral, and analytic measures based on mathematical tasks and Wolfram Alpha-supported analysis. The study adopted a quantitative predictive research design using a purposive sample of 72 students in their last year of studying mathematics teaching within a single institution (195 task-level observations). The dependent variable, preparation level, was conceptualized at the individual level by utilizing a preparation rubric of the institution and categorized into three tertile groups: low, medium, and high. Given the hierarchical structure of the data, statistical analyses were performed using linear mixed-effects models with Student_ID as a random effect. In order to avoid leakage of information among repeated tasks from the same candidate, student-level features were extracted from the task-level observations in the predictive modeling phase. Various machine learning models were tested using stratified five-fold cross-validation with hyperparameter optimization. The findings indicated statistically significant variations between the preparation levels in terms of error rate, symbolic steps of solutions, test scores, completion times, and progress rates. The Random Forest model yielded the highest performance, with accuracy of 0.889 ± 0.064 and macro-F1 score of 0.888 ± 0.064. These results indicate that integrating features based on academic and behavioral characteristics, along with analytical features, yields an efficient approach to modeling teacher preparation level. However, it should be noted that further research and validation are required on larger datasets from different institutions.