Markov Blanket Feature Selection Using Representative Sets.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Markov blankets in Bayesian networks are widely used for optimal feature selection.
• Uniqueness of Markov blankets is guaranteed only if the data distribution faithfully represents the Bayesian network.
• Violations of the faithful condition can lead to non-unique Markov blankets, posing challenges for feature selection.

Purpose of the Study:

• To address the non-uniqueness issue of Markov blankets when the faithful condition is violated.
• To propose a novel concept of representative sets for robust Markov blanket feature selection.
• To introduce the selection via group alpha-investing (SGAI) algorithm for classification tasks.

Main Methods:

• Development of the concept of representative sets.
• Design of the selection via group alpha-investing (SGAI) algorithm.
• Empirical evaluation using a comprehensive set of real-world datasets.

Main Results:

• The proposed SGAI algorithm effectively performs Markov blanket feature selection using representative sets.
• Empirical studies demonstrate that SGAI outperforms existing state-of-the-art Markov blanket feature selectors.
• SGAI also shows superior performance compared to other well-established feature selection methods.

Conclusions:

• The SGAI algorithm provides a robust solution for Markov blanket feature selection, even when data violates the faithful condition.
• Representative sets enhance the reliability of Markov blanket identification.
• SGAI offers a promising approach for improving classification performance through effective feature selection.