Empirical analysis of rough set categorical clustering techniques based on rough purity and value set

Clustering a set of objects into homogeneous groups is a fundamental operation in data mining. Recently, attention has been put on categorical data clustering, where data objects are made up of non-numerical attributes. The implementation of several existing categorical clustering techniques is challenging as some are unable to handle uncertainty and others have stability issues. In the process of dealing with categorical data and handling uncertainty, the rough set theory has become well-established mechanism in a wide variety of applications including databases. The recent techniques such as Information-Theoretic Dependency Roughness (ITDR), Maximum Dependency Attribute (MDA) and Maximum Significance Attribute (MSA) outperformed their predecessor approaches like Bi-Clustering (BC), Total Roughness (TR), Min-Min Roughness (MMR), and standard-deviation roughness (SDR). This work explores the limitations and issues of ITDR, MDA and MSA techniques on data sets where these techniques fails to select or faces difficulty in selecting their best clustering attribute. Accordingly, two alternative techniques named Rough Purity Approach (RPA) and Maximum Value Attribute (MVA) are proposed. The novelty of both proposed approaches is that, the RPA presents a new uncertainty definition based on purity of rough relational data base whereas, the MVA unlike other rough set theory techniques uses the domain knowledge such as value set combined with number of clusters (NoC). To show the significance, mathematical and theoretical basis for proposed approaches, several propositions are illustrated. Moreover, the recent rough categorical techniques like MDA, MSA, ITDR and classical clustering technique like simple K-mean are used for comparison and the results are presented in tabular and graphical forms. For experiments, data sets from previously utilized research cases, a real supply base management (SBM) data set and UCI repository are utilized. The results reveal significant improvement by proposed techniques for categorical clustering in terms of purity (21%), entropy (9%), accuracy (16%), rough accuracy (11%), iterations (99%) and time (93%). vii

Name dc:type.qualificationname

phd

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

Universiti Tun Hussein Onn Malaysia

Year dc:date.issued

2017

Author dc:creator

• Uddin, Jamal

• QA76 Computer software

Language dc:language

en