A Novel Approach for Intrinsic Dimension Estimation

Dimensionality reduction approaches are crucial in various applications of machine learning tasks such as computer vision, robotics, natural language processing, medical diagnosis, recommendation systems or industrial IoT applications such as predictive maintenance which need to generate and process large amounts of data and variables. In general, dimensionality reduction improves the performance of machine learning tasks' by removing redundant features. In this regard, both linear and non-linear dimensionality reduction methods, specifically the manifold learning techniques are particularly efficient since they are based on the preservation of the geometric structure of the original feature space. In this manner, there are several approaches already available and studied extensively in the literature such as principal component analysis (PCA), Multidimensional scaling (MDS), Laplacian Eigenmaps (LE) and other. We refer the reader to (Lee and Verleysen, 2007) for a comprehensive survey of the available methods.