Graph Embedding Discriminative Unsupervised Dimensionality Reduction

Abstract

In this thesis, a novel graph embedding unsupervised dimensionality reduction method was proposed. Simultaneously, we assigned the adaptive and optimal neighbors on the basis of the projected local distances, thus we developed the dimensionality reduction along with the graph construction. The clustering results could be directly exhibited from the learnt graph which has the explicit block diagonal structure.The analysis of experimental result on different databases also determines that the proposed dimensionality reduction method is superior to other related dimensionality reduction methods, like PCA and LPP. In this study, we use synthetic data and real-world benchmark data sets. Also experimental results from the clustering experiments revealed the proposed dimensionality reduction method outperformed other clustering methods, such as K-means, Ratio Cut, Normalized Cut and NMF.