Measuring Data Quality: Predictive Accuracy vs. Similarity of Decision Trees

Abstract

<p>Nowadays huge amount of data is regularly being collected for various purposes by organizations and companies, including business, government departments, and medical service providers. Data mining techniques are often used on these gigantic datasets to discover previously hidden information. Upon release of these datasets for data mining, individual sensitive and delicate information is at high risk of being exposed to unauthorised disclosure. Due to the growing public concern about privacy, many control techniques have been proposed to protect confidentiality of individual information. Some of these techniques involve perturbing datasets by adding a noise to data in some controlled fashion. The effectiveness of such techniques is typically evaluated by measuring the security and data quality of perturbed dataset. In this paper we experimentally evaluate the data quality by comparing the prediction capability of decision trees and neural networks, built from original and perturbed datasets. We then compare this evaluation technique to the one that uses logic rules associated with the decision tree classifiers. </p>