Design And Analysis Of Entropy Based Ranking Measures For Knowledge Discovery

Abstract

With the widespread use of digital data in this digital era, many features of data as knowledge are extensively stored for providing visual information. However, in large databases the storing of data exists in multiple forms like computational, statistical, and mathematical etc. Therefore, to provide information from the complete data requires ample storage space, high execution time and cost for pro cessing. Hence, there is a need to create a system that supports easy and quick finding new knowledge from databases. This process is known as knowledge dis covery. In this research work, the various novel designs and analysis of entropy based ranking measures are proposed and investigated for knowledge discovery. This research work is broadly classified in three parts, first to design feature se lection algorithms using entropy that select informative features, second to design data reduction algorithms using information and fuzzy approaches and third, to develop new algorithms with the use of evolutionary algorithms for knowledge discovery. The proposed feature selection algorithms based on fuzzy entropy for removal of redundant and irrelevant attributes results in reducing the dataset size and in turn takes lesser time for computation, more understandable and intel ligible. Thus, the attribute reduction proved to be an efficient tool for feature selection using fuzzy entropy with the proper choice of filter based on fuzzy uncer tainty, roughness, and decomposition as it helps in efficient transformation of data to knowledge. The feature selection of binary class data sets using hybrid entropy, rough entropy and SVD entropy are examined and evaluated on benchmark data sets of different fields. The presented algorithms provide non-dominated solutions to obtain a smaller number of features and superior classification performance on different classifiers such as k-Nearest Neighbours (KNN), Support Vector Machine (SVM), and Multi-Layer Perceptron - Artificial Neural Network (MLP-ANN). The performance of ...