Using multivariate random forests for predicting learning trajectories from digital training data [Author Accepted Manuscript]

Heterogeneous treatment effects allow tailoring of interventions. In personalized education, computerized systems integrate assessment and training in a data-rich environment. Interventions are often adapted by predicting developmental trajectories and assigning optimal treatment regimes. While traditional multivariate methods focus on explanation and inference, they are suboptimal for prediction. This study explores whether the multivariate random forest (MRF) outperforms univariate models in predicting trajectories in an educational training system for dyscalculia. Using principal component analysis and nested cross-validation, MRF was compared with univariate random forests and multivariate lasso regression. Results show MRF performs comparably but is not superior. A thorough initial assessment remains a strong predictor of non-linear learning trajectories, highlighting the value of psychometric testing and interpretable feature engineering.