Visualization Methods for Exploratory Data Analysis

We investigate, in this paper, the use of linear and nonlinear methods for low-dimensional visualization. Results of four projection methods of different categories: PCA, KPCA, Sammon, and CCA are tested and compared on synthetic and real datasets. In order to evaluate the structure preservation quality of projected data, dx-dy criterion is usually applied which permits to visualize distortions between original data points and their projections. However, in the context of exploratory data analysis, the appropriate projection method is that which reveals the presence of natural clusters inside the scatter plots. So, we propose to use the rate accuracy of K-means clustering algorithm applied on projected data to compare the quality of projection methods