Subspace-Based Aggregation for Enhancing Utility, Information Measures, and Cluster Identification in Privacy Preserved Data Mining on High-Dimensional Continuous Data

Abstract

Clustering is a data mining technique that has been effectively used in the last few decades for knowledge extraction. Privacy is a major problem while releasing data for clustering and therefore privacy-preserving data mining (PPDM) algorithms have been developed. Aggregation is a popular PPDM technique that has been used. However, in the last few years, certain applications require that data mining be performed on high-dimensional data. The present privacy preservation techniques perform aggregation in a univariate manner along each dimension. This affects the utility measures, information measures, and especially retention of original clusters. This paper proposes a new technique called as subspace-based aggregation (SBA). SBA categorizes the dimensions into dense and non-dense subspaces based on the density of points. Aggregation is performed separately for dense and non-dense subspaces. This approach helps to maximize utility measures, information measures, and retention of clusters. SBA is run on high-dimensional continuous datasets from UCI Machine Learning repository. SBA is compared with related work methods such as SINGLE, SIMPLE, MDAV, and PPPCA. SBA provides an improvement of 66% in utility, 400% in cluster identification, 5% in co-variance, and standard deviation.