Structure Learning of Bayesian Network Based on Adaptive Thresholding.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Traditional methods for identifying dependencies in Bayesian network classifiers (BNCs) are computationally expensive, especially for complex structures.
• Existing filter and wrapper approaches identify non-significant dependencies individually, leading to inefficiency.
• Overfitting and structural unreliability are common issues in BNCs due to inefficient dependency selection.

Purpose of the Study:

• To investigate novel methods for efficient dependency identification in k-dependence Bayesian classifiers.
• To enhance the reliability and reduce overfitting of BNC structures.
• To improve classification performance by effectively filtering redundant dependencies.

Main Methods:

• Developed MI-based feature selection and CMI-based dependence selection techniques.
• Implemented a novel adaptive thresholding method to filter redundant dependencies.
• Evaluated the proposed methods on 30 datasets from the UCI machine learning repository.

Main Results:

• Adaptive thresholding effectively distinguishes between dependencies and independencies.
• The proposed algorithm demonstrates competitive classification performance against state-of-the-art BNCs.
• The methods achieved favorable results in terms of 0-1 loss, root mean squared error, bias, and variance.

Conclusions:

• The proposed adaptive thresholding approach offers an efficient and effective solution for dependency selection in BNCs.
• This method enhances BNC structure reliability and mitigates overfitting.
• The techniques provide a valuable advancement for Bayesian network classifier development and application.