Privacy-preserving Data Mining for Personalized Marketing

Abstract

In electronic commerce markets, firms can easily achieve customers’ personal information such as identity, demographic information, and shopping behavior. Researches in securing statistical database had introduced several tools and methods to secure such statistical database with sensitive personal information. Although the data perturbation methods secure the database very effectively, it is not applicable to the application beyond the simple statistical analysis on means, variances and covariance. In this paper, I suggest Multiple Staged Slice Perturbation Methods in order to apply them to RFM analysis. My study shows the possibility of applying a simple modification to perturbation methods in order to be able to perform the RFM analysis. My method of slicing the database into decile and perturbing each decile separately would maintain the mean and the standard deviation of each decile. I showed that current data security methods may not be applicable to some business analysis that deals with more than the mean, standard deviation and covariance between variables. Since perturbation method guarantees protection against exact disclosure, there is no threat of exact disclosure even if data is partitioned into small pieces and perturbed individually. However, because partitioning limits range of shuffling effect, partial disclosure is possible. Therefore, for achieving the maximum utility while preserving maximum security level, the number of partition should be minimized.