Mixed Bayesian networks: a mixture of Gaussian distributions

Area of Science:

• Statistics
• Machine Learning
• Probabilistic Graphical Models

Background:

• Mixed Bayesian networks model complex systems with discrete and continuous variables.
• Accurate estimation of continuous variable density functions is crucial for network analysis.
• Existing methods may face challenges in learning these densities from sample data.

Purpose of the Study:

• To propose a comprehensive method for estimating density functions of continuous variables within mixed Bayesian networks.
• To leverage graph structures and sample data for accurate density estimation.
• To adapt existing algorithms for efficient estimation and inference.

Main Methods:

• Approximating continuous variable densities using Gaussian mixture models.
• Employing a stochastic Expectation-Maximization (EM) algorithm for density estimation.
• Utilizing a modified junction tree algorithm for probabilistic inference.

Main Results:

• The proposed method effectively learns the shape of densities from continuous variable samples.
• Gaussian mixture models provide a suitable approximation for the true distributions.
• The stochastic EM algorithm and adapted inference method demonstrate computational feasibility.
• Simulated pharmacokinetic data showed satisfactory fitting of true distributions.

Conclusions:

• The developed method offers a robust approach for density estimation in mixed Bayesian networks.
• The technique is suitable for practical applications, particularly in pharmacokinetic modeling.
• This work contributes to the advancement of probabilistic graphical model analysis for continuous variables.