Abstract
<p>Recent advances in assessments have challenged traditional psychometric methods in analyzing high-dimensional sparse data and achieving accurate predictions. Meanwhile, machine learning methods, such as Matrix Factorization (MF) sharing similarities with the common factor model, show promising potential for addressing those challenges but lack interpretability. To fill this gap, we propose Confirmatory Matrix Factorization (CMF) as a computational psychometric method based on MF and compare it to Confirmatory Factor Analysis (CFA) and Confirmatory Penalized Factor Analysis (CPFA). The comparison is based on 48,600 data sets generated by the common factor model under various characteristics, alongside two real data sets. The results show that CFA and CPFA struggles with non-convergence in high-dimensional sparse data, whereas CMF performs effectively with high computation efficiency. CMF generally outperforms CFA and CPFA regarding prediction performance in most cases. CMF can successfully recover factor loadings and the estimated factor scores from three methods are comparable. Additionally, missing values negatively affect the performance of the three methods. Overall, CMF can be an interpretable and robust alternative to CFA and CPFA, particularly in emerging assessments where CFA and CPFA is less effective.</p>