Novel clustering algorithms: entropy based neighborhood merging (ENM) and simultaneous feature selective clustering (SFSC)

Clustering is an unsupervised machine learning approach that is frequently used by researchers in data science. It can be defined as the process of dividing a data set into subsets such that similar elements are in the same subset and/or dissimilar elements are in different subsets. In the state of art for clustering literature, clustering methodologies are challenged by some issues such as arbitrary geometric shaped clusters, density variations, multidimensional feature space and overcoming these issues within reasonable computational complexity and within a user friendly environment. In this thesis, novel clustering methodologies (Entropy Based Neighborhood Merging – ENM and Simultaneous Feature Selective Clustering - SFSC), are proposed for arbitrary geometric shaped clusters along with the density variations and multidimensional feature space respectively. These algorithms are based on some statistical concepts such as Shannon's Entropy, Symmetric Relative Entropy and Isotropic Position. Experimental analysis of both ENM and SFSC are carried out on benchmark datasets to show their applicability and efficiency on clustering. Furthermore, two case studies are given (seismic zone detection and churn analysis) for the real-time applications of both algorithms. The experimental analyzes and the case studies showed that, ENM and SFSC are efficient methodologies for the relevant challenging issues and they have high levels of applicability, interpretability and user friendly character.