Performance evaluation of non-linear techniques UMAP and t-SNE for data in higher dimensional topological space

Abstract

Dimension reduction is the vital area in data science & analytics for visualization, and significant pre-processing step for artificial intelligence and machine learning based analysis. For 3D visualization and data analytics of higher dimensional data, it is mandatory to reduce it into lower dimensional subspace. Higher dimensional data existence is everywhere in all type of sectors like Telecom, healthcare infrastructure, Finance, Banking, Transport, eCommerce etc. Applying regression analysis directly on higher dimensional data in machine learning or AI based analytics not recommended. Generally, before analysis, such data is reduced to lower dimensional topological subspace, maintaining the essence of original data. In this paper, a performance comparison of two competitive projection-based non-linear dimension reduction techniques - UMAP and t-SNE with a combination of PCA as a linear based method is analyzed with telecom gateway data. Apart from this, both non-linear techniques are compared based on 3D visualization of handwritten digits images.