Title: Fuzzy clustering-based non-linear dimensionality reduction
Authors: Mika Sato-Ilic - University of Tsukuba (Japan) [presenting]
Abstract: Non-linear dimensionality reduction is a key component of recent machine learning based techniques for large and complex data. This is because such data exist in a mathematically non-linear observational space, and it is difficult to obtain a meaningful and clear latent structure of the observational data with conventional linear dimensional reduction methods. Foremost reasons for this difficulty are the sensitivity of the data in the non-linear observational space and that the simple linear dimensionality reduction techniques lose important sensitive information which is part of the original observation. We propose a new fuzzy clustering-based non-linear dimensionality reduction method to overcome these problems. This method utilizes fuzzy clustering to consider the nonlinearity of the observational space and involves captured features of the nonlinearity to the conventional linear dimensionality reduction method. In addition, by including the difference of fuzzy clustering results over the times from the originally observed data over the times, visualization for the difference of data over times by using the proposed method is presented. Several numerical examples show the better performance of the proposed method.