Positive region: An enhancement of partitioning attribute based rough set for categorical data

Abstract

Datasets containing multi-value attributes are often involved in several domains, like pattern recognition, machine learning and data mining. Data partition is required in such cases. Partitioning attributes is the clustering process for the whole data set which is specified for further processing. Recently, there are already existing prominent rough set-based approaches available for group objects and for handling uncertainty data that use indiscernibility attribute and mean roughness measure to perform attribute partitioning. Nevertheless, most of the partitioning attribute methods for selecting partitioning attribute algorithm for categorical data in clustering datasets are incapable of optimal partitioning. This indiscernibility and mean roughness measures, however, require the calculation of the lower approximation, which has less accuracy and it is an expensive task to compute. This reduces the growth of the set of attributes and neglects the data found within the boundary region. This paper presents a new concept called the "Positive Region Based Mean Dependency (PRD)”, that calculates the attribute dependency. In order to determine the mean dependency of the attributes, that is acceptable for categorical datasets, using a positive region-based mean dependency measure, PRD defines the method. By avoiding the lower approximation, PRD is an optimal substitute for the conventional dependency measure in partitioning attribute selection. Contrary to traditional RST partitioning methods, the proposed method can be employed as a measure of data output uncertainty and as a tailback for larger and multiple data clustering. The performance of the method presented is evaluated and compared with the algorithmes of Information-Theoretical Dependence Roughness (ITDR) and Maximum Indiscernible Attribute (MIA).