Feature selection via Class-specific Approximate Markov Blanket and Rough Set-based Mapping

Abstract

Markov Blanket (MB) is a currently popular approach to feature selection that helps to effectively select correlated features and eliminate redundant features. However, existing MB-based approaches involve complex computations and extensive search. Therefore, we propose a novel concept, Class-specific Approximate Markov Blanket (CSAMB), to solve the above two problems from a class-specific perspective. This concept involves the transformation of decision attributes and features in the specific class using a proposed Rough Set-based Mapping (RSM) method, facilitating the selection results with high classification correlation and low inter-redundancy. The RSM not only preserves the positive, negative and boundary regions of a specific class with respect to a given feature, but also accurately quantifies the relationship between features within that class. Notably, we explore the approximate upper and lower bounds of grouping of correlation features via CSAMB. We then design a CSAMB-based algorithm, and extend it to two variants: CSAMB-min and CSAMB-max using the approximate upper and lower bounds, which demonstrates the performance range of our algorithm. Experiments shows that our algorithms outperform state-of-the-art algorithms regarding accuracy and efficiency, especially for large-scale and high-dimensional datasets.