DADR: Deep adaptive dimensionality reduction architecture based on the coefficient of variation

Abstract

The traditional deep dimensionality reduction methods corresponding to the depth of the model and the dimensionality of each layer depend on empirical settings. In this paper, we propose a deep adaptive dimensionality reduction architecture (DADR) based on the coefficient of variation and Gaussian restricted Boltzmann machine (GRBM) for achieving adaptivity of depth and dimensionality in the dimensionality reduction process. To verily the validity of the proposed model, we introduce two unsupervised algorithms, K-means and spectral clustering (SC), to compare the DADR architecture with all original features, shallow GRBM model, PCA and two advanced feature selection-based dimensionality reduction algorithms (CNAFS and UFSwithOL), respectively. The final experimental results show the performance of the proposed DADR architecture is demonstrated to be superior to the other algorithmic models. The source code is available at https://github.com/dingxm99/DADR.