Discriminative Structure Learning of Bayesian Network Classifiers from Training Dataset and Testing Instance.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• The increasing volume of data necessitates scalable Bayesian networks for improved classification and expressivity.
• Existing k-dependence Bayesian classifiers (KDB) use efficient but sub-optimal attribute sorting, neglecting conditional dependencies.

Purpose of the Study:

• To propose a novel sorting strategy and extend KDB to unrestricted ensemble networks (UKDB) for enhanced classification.
• To improve the handling of conditional dependencies and attribute relationships in Bayesian networks.

Main Methods:

• Developed a novel attribute sorting strategy based on Markov blanket analysis and target learning.
• Introduced UKDB P and UKDB T, ensemble networks modeling different data spaces for complementary classification.
• Utilized a target learning framework where each test instance builds a specific Bayesian model.

Main Results:

• The proposed UKDB approach demonstrated effectiveness and robustness across 11 datasets, including the Wisconsin breast cancer database.
• UKDB outperformed simpler models like Naive Bayes and Tree Augmented Naive Bayes in classification tasks.
• The novel sorting strategy effectively captures conditional dependencies, improving upon KDB.

Conclusions:

• UKDB offers a significant advancement in scalable Bayesian network classification, particularly for large and complex datasets.
• The ensemble approach and novel sorting strategy enhance both performance and the ability to model intricate data relationships.
• This research provides a more expressive and accurate method for Bayesian network classification.