Variational Bayes Inference Algorithm for the Saturated Diagnostic Classification Model

Area of Science:

• Educational Measurement
• Psychometrics
• Computational Statistics

Background:

• Saturated diagnostic classification models (DCM) offer flexible attribute mastery diagnosis.
• Existing saturated DCM formulations hinder the derivation of conditionally conjugate priors for variational Bayes (VB) inference.

Purpose of the Study:

• To propose a novel mixture formulation for saturated DCM.
• To develop a scalable and computationally efficient VB inference algorithm for saturated DCM.

Main Methods:

• Introduced a novel mixture formulation of saturated DCM.
• Developed a VB inference algorithm based on the new formulation.
• Conducted simulation studies and analyzed a real educational dataset.

Main Results:

• The proposed VB algorithm enables scalable and efficient Bayesian estimation.
• Simulation studies confirmed parameter recovery across various conditions.
• The method is well-suited for sequentially available data, like in computerized diagnostic testing.
• VB inference was significantly faster than Markov Chain Monte Carlo (MCMC) with comparable estimates on real data.

Conclusions:

• The novel mixture formulation and VB algorithm provide a practical solution for computational challenges in saturated DCM.
• This approach enhances the efficiency and scalability of Bayesian estimation for diagnostic classification models.