Network-based dimensionality reduction and analysis

Abstract

Network analysis opens new horizons for data analysis methods, as the results of ever-developing network science can be integrated into classical data analysis techniques. This paper presents the generalized network-based dimensional analysis (GNDA) method. The main contributions of this paper are as follows: (1) The proposed GNDA method handles high dimensional low sample size datasets. In addition, compared with existing methods, we show that only the proposed GNDA method adequately estimates the number of latent variables. (2) The proposed GNDA already considers any symmetric and nonsymmetric similarity functions between indicators (i.e., variables or observations) to specify latent variables. The proposed GNDA method is compared with traditional dimensionality reduction methods on various simulated and real-world datasets. The implementation of the proposed method can be downloaded from the official CRAN site.