Accelerating item factor analysis on GPU with Python package xifa.

Area of Science:

• Psychometrics and Statistical Modeling
• Computational Statistics
• Machine Learning Applications in Social Sciences

Background:

• Item factor analysis (IFA) is essential for understanding measurement properties.
• Marginal maximum likelihood (MML) is the standard frequentist estimation method for IFA.
• Fitting high-dimensional IFA models with MML presents significant computational challenges.

Purpose of the Study:

• To demonstrate the reduction in computational time for MML in large-scale IFA using GPU acceleration.
• To introduce and evaluate a new Python package, xifa, for accelerated item factor analysis.
• To compare the performance of the proposed GPU-accelerated method against existing computational approaches.

Main Methods:

• Development of a Python package, xifa, implementing a vectorized Metropolis-Hastings Robbins-Monro (VMHRM) algorithm.
• Leveraging graphics processing unit (GPU) computing for parallel processing of IFA models.
• Numerical experiments comparing VMHRM on GPU with CPU-based methods and other algorithms.

Main Results:

• The GPU-implemented VMHRM algorithm achieved a 33-fold speedup compared to its CPU version.
• VMHRM on GPU significantly outperformed Bock-Aitkin EM, mirt's MHRM (CPU), and importance-weighted autoencoders (GPU) for models with five or more factors.
• The GPU-accelerated VMHRM is highly suitable for high-dimensional IFA with large datasets.

Conclusions:

• GPU computing offers a powerful solution for accelerating computationally intensive psychometric modeling tasks like high-dimensional IFA.
• The xifa package provides an efficient tool for researchers dealing with large-scale IFA.
• GPU acceleration is poised to become a central component in future large-scale psychometric modeling research.