An Entropy-based Data Reduction Method for Data Preprocessing

Abstract

The primary task in data mining is to find potential patterns or to discover hidden and useful knowledge from given data sets. However, with the increasing data quantity and exploding complexity, the capabilities of dealing with massive data becomes very crucial. The data preprocessing module is an integral part of data mining procedure, which aims to optimize input data usability for subsequent tasks such as classification, clustering, association analysis as well as other data mining algorithms. In general, data preprocessing procedures can effectively reduce the computational complexity while as possible to ensure accuracy and efficiency of prediction or classification, but meanwhile it even can assist to extract some unknown knowledge before applying more advanced data mining algorithms. This research proposes a three-patterns feature variables technique and an entropy-based data reduction (EBDR) algorithm for data preprocessing based on information theory. The goal is to explore high-purity subsets in which the values of an attribute are directly linked to specific class labels. The results of experiments demonstrate the efficacy of EBDR algorithm on datasets of varying sizes.