TANGENT SPACE BASED DIMENSIONALITY REDUCTION TECHNIQUE

Abstract

Many Pattern Recognition and machine learning tasks, such as clustering and classification, may become challenging due to high dimensionality of data. Data with high dimensions are usually noisy and redundant, thus the performance of subsequent procedures is degraded. So, Dimensional Reduction methods are required to give better results of classification or clustering tasks. As a result lot of supervised/unsupervised dimensionality reduction methods are implemented. Among them, most unsupervised DR methods focus on local geometrical structure/information while reducing the dimension of data, but almost all the methods neglect the fact that the dimension of local data is higher than the number of local data points (neighbors) which exhibits ”curse of Dimensionality” problem. It will impact on the performance of methods. So, our main focus here is to solve the above problem occurring while capturing local information. In the proposed method, we have used kernel principal component analysis to solve the above problem. It will preserve a considerable amount of information if data is non-linear in nature and tangent space is used to preserve the local structure. To check its effectiveness, we estimated it on various standard datasets such as Ecoli, USPS-handwritten digit dataset, face94 and Olivetti face image dataset and giving better results.