Data reduction approach for sensitive associative classification rule hiding

Abstract

When a business unit shares data with another unit, there could be some sensitive patterns which should not be disclosed. In order to remove or hide a sensitive pattern in data sharing scenario, the data set needs to be modified such that the sensitive pattern becomes uninteresting according to the pre-specified interestingness threshold (s). However, data quality of the given data set should also be maintained, otherwise, the sharing will be meaningless. Existing data modification algorithms usually use data perturbation approach, i.e. changing some data values in a given data set from an original value to another value. Though, it could hide sensitive patterns and maintain data quality, such the approach could not be applied in a situation where real data are required. In this paper, we explore an alternate approach for sensitive pattern hiding problem, data reduction, i.e. removing the whole selected tuples. By data reduction, every tuple in modified data sets is real data without any change. The focused pattern type is associative classification rule. The impact on data quality is denoted as the numbers of false-dropped rules and ghost rules. The experiments are conducted to evaluate the approach and the results have shown that data reduction approach can produce data sets with high data quality, thus it is applicable to the problem. © 2008, Australian Computer Society, Inc.