n-means: Adaptive Clustering Microaggregation of Categorical Medical Data

Abstract

Huge amount of information is managed and shared publically by the individuals and data controllers. Publically shared data contains information that can reveal identity of users, thus affecting privacy of individuals. To palliate these disclosure risks, Statistical Disclosure Control (SDC) methods are applied to the data before it is released. Microaggregation is one of the SDC methods that aggregate similar records into clusters, and then transform them into m indistinguishable records. K-means is a famous data mining clustering algorithm for continuous data, which iteratively maps similar elements into k-cluster until they all converge. However, adapting k-means algorithm for categorical multivariate is a challenging task due to high dimensionality of attributes. In this paper, we extend k-means clustering algorithm to achieve notion of microaggregation of structured data. Moreover, to preserve data utility, we extend fixed clustering nature of this algorithm to adaptive size clusters. For this purpose, we introduce n-means clustering approach that construct clusters based on the semantics of the datasets. In experiments, we proved significance of our proposed system by measuring cohesion of clusters and information loss for utility purpose.