Privacy Preserving Feature Selection via Voted Wrapper Method for Horizontally Distributed Medical Data

Abstract

Feature selection plays a crucial step for data mining algorithms via eliminating the curse of dimensionality. Many feature selection approaches are developed for analyzing centralized data on the same location. In recent years, multi-source biomedical data mining methods have been developed to analyze different distributed databases at different locations such as different hospitals. However, a major concern is privacy of sensitive personal medical records in different hospitals. Therefore, as the needs for new privacy preserving distributed data mining algorithms increase, it is necessary to develop new privacy preserving feature selection algorithms for biomedical data mining. In this paper, a privacy preserving feature selection method named “Privacy Preserving Feature Selection algorithm via Voted Wrapper methods (PPFSVW)” is developed. This method was tested on six benchmark datasets under two testing scenarios. Our experimental results indicate that the proposed algorithm workflow can work effectively to improve the classification performance regarding accuracy via selecting informative features and genes. Besides, the proposed method can make the classifier achieve higher or same level classification accuracy with fewer features compared with those sophisticated methods, such as SVM-RFE, RSVM and SVM-t. More importantly, the individual private information can be protected during the whole feature selection process.