On addressing accuracy concerns in privacy preserving association rule mining

Abstract

Randomized Response techniques have been empirically investigated in privacy preserving association rule mining. In this paper, we investigate the accuracy (in terms of bias and variance of estimates) of both support and confidence estimates of association rules derived from the randomized data. We demonstrate that providing confidence on data mining results from randomized data is significant to data miners. We propose the novel idea of using interquantile range to bound those estimates derived from the randomized market basket data. The performance is evaluated using both representative real and synthetic data sets. © 2008 Springer-Verlag Berlin Heidelberg.