Clustering Using Local and Global Exponential Discriminant Regularization

In recently reported clustering approaches, both local and global information were utilized in order to effectively learn nonlinear manifold in image dataset. However, in each of these clustering approaches, regularization parameter had to be included to handle small-sample-size (SSS) problem of linear discriminant analysis (LDA). Due to which, we have to optimize a number of clustering parameters to report optimal clustering performance in these clustering models. In this study, we propose less-parameterized Local and Global Exponential Discriminant Regularization (LGEDR) clustering model. Our proposed LGEDR model is based on exponential discriminant analysis (EDA) in which SSS problem of LDA is handled without including regularization parameter. Because, no discriminant information of LDA is lost in EDA, clustering performance of the proposed LGEDR model is comparable over existing state-of-art clustering approaches on 12 benchmark image datasets. Further, due to less-parameterized nature, proposed LGEDR model is computationally efficient over existing clustering approaches that utilized both local and global information in image data.