A Deep Learning Algorithm for High-Dimensional Exploratory Item Factor Analysis.

Area of Science:

• Psychometrics
• Computational Statistics
• Machine Learning

Background:

• Marginal maximum likelihood (MML) estimation is standard for item response theory (IRT) models.
• Current MML methods (e.g., MH-RM, VI) are computationally intensive for large datasets and numerous latent factors.

Purpose of the Study:

• To introduce and evaluate a deep learning-based variational inference (VI) algorithm for exploratory item factor analysis (IFA).
• To assess the computational speed and estimation accuracy of the proposed deep learning approach for large-scale psychometric modeling.

Main Methods:

• Utilized an importance-weighted autoencoder (IWAE), a deep artificial neural network, to approximate the MML estimator in exploratory IFA.
• Employed an importance sampling technique within the IWAE framework, adjusting the number of importance-weighted (IW) samples to balance approximation accuracy and computational efficiency.

Main Results:

• The IWAE approach demonstrated computational speed advantages over traditional MML methods for large datasets and multiple latent factors.
• Simulation studies indicated that increasing sample size or IW samples improved IWAE estimation accuracy, with results comparable to MH-RM but in less time.
• The IWAE showed comparable or superior speed to constrained joint maximum likelihood estimation.

Conclusions:

• The deep learning-based IWAE provides a computationally efficient and accurate method for fitting exploratory IFA models, particularly for large-scale data.
• This approach offers a promising alternative to existing time-consuming MML estimation procedures in psychometrics.