Abstract
Feature selection is an important process in data analysis and data mining. The increasing size, complexity, and multi-valued nature of data necessitate the use of Symbolic Data Analysis (SDA), which utilizes symbolic objects instead of classical tables, for data analysis. The symbolic objects are created by using abstraction or generalization techniques on individuals. They are a representation of concepts or clusters. To improve the description of these objects, and to eliminate incoherencies and over-generalization, using dependencies between variables is crucial in SDA. This study shows how correlation dependencies between variables can be processed on Boolean Symbolic Objects (BSOs) in feature selection. A new feature selection criterion that considers the dependencies between variables, and a method of dealing with computation complexity is also presented.