Using Diagnostic Classification Models to Validate Attribute Hierarchies and Evaluate Model Fit in Bayesian Networks

Area of Science:

• Computational neuroscience
• Machine learning
• Statistical modeling

Background:

• Bayesian inference networks (BayesNets) and diagnostic classification models (DCMs) are both used for modeling complex systems.
• Existing research has not fully elucidated the relationship and potential equivalencies between these two modeling approaches.

Purpose of the Study:

• To investigate and demonstrate the parameterization equivalency between BayesNets and DCMs.
• To propose a novel model-comparison framework for assessing the fit of BayesNets.
• To establish the nested relationships between BayesNets, saturated DCMs, and Hierarchical DCMs under specific conditions.

Main Methods:

• Empirical examination of parameterization equivalencies between BayesNets and DCMs.
• Development of a model-comparison framework for BayesNets.
• Demonstration of BayesNets nested within saturated DCMs.
• Analysis of Hierarchical DCMs nested within BayesNets and saturated DCMs for linear attribute hierarchies.

Main Results:

• Established equivalency in parameterizations between BayesNets and DCMs.
• Demonstrated that BayesNets are nested within saturated DCM structural models.
• Showed that Hierarchical DCMs are nested within BayesNets and saturated DCMs when attributes exhibit a linear hierarchy.
• Validated the proposed framework and model-fit testing strategy using simulated and empirical data.

Conclusions:

• BayesNets and DCMs share fundamental equivalencies in their parameterizations.
• The proposed model-comparison framework provides a robust method for evaluating BayesNets.
• Understanding the nested relationships aids in selecting appropriate models for diagnostic classification tasks.