Simultaneous Clustering and Dynamic Feature Selection Methodology Based on Dimensional Isotropy Preservation

Abstract

There are some challenging issues in clustering problems and clustering methodologies, which frequently attract researchers' interests. Ability of handling high dimensional feature space is among them, since robustness of all machine learning algorithms is reduced by the concept of curse of dimensionality. In unsupervised machine learning literature, a general consensus can be seen on the clustering process: If the data is high dimensional, as a preprocessing step, first reduce the number of dimension using dimension reduction techniques, then cluster data. However, such a kind of cascaded way may cause information loss and redundant repeating operations, consequently reduce the validity of the clustering output and may yield low quality clustering schema and bring more computational burden. In the literature there is no methodology, which is based on classical statistical concept and capable of handling high dimensional data, except graph theory based ones. In that study; a hybrid approach is proposed for high dimensional data clustering which is mainly based on subspace clustering and projected clustering using a statistical concept: isotropic position. More precisely, proposed algorithm deals with clustering by simultaneously feature selection upon dimensional isotropic preservation. As a result, preprocessing step for dimensional reduction will not be required, when proposed methodology is preferred for the clustering process. Moreover, since the principal features are selected dynamically in the proposed methodology by not a holistic concern of the dataset; high quality clustering schema is expected, while low computational burden will be incurred. Experimental analysis of the proposed algorithm points out promising results for simultaneous clustering and dimension reduction.