Pinball twin bounded support vector clustering

Abstract

Clustering is a very popular approach in machine learning for unlabelled data. Twin support vector clustering (TWSVC) and the twin bounded support vector clustering (TBSVC), plane-based clustering algorithms introduced recently, work on twin support vector machine (TWSVM) principles and are used in widespread clustering problems. However, both TWSVC and TBSVC are sensitive to noise and su↵ers from low re sampling of data stability due to the use of hinge loss. The pinball loss features noise insensitivity and stability for re-sampling of data. Within this thesis, we first present basic formulations of the previous methods in plane based clustering and discuss their shortcomings. Then we propose two plane-based clustering methods, twin bounded support vector clustering using pinball loss (pinTBSVC) and sparse twin bounded sup port vector clustering using pinball loss (pinSTBSVC) which inherits various attributes from previous plane based clustering algorithms. Sparse solutions help to create better generalized solutions in the clustering problems; hence we attempt to use maximum margin regularization term to propose pinSTBSVC. The proposed pinTBSVC and pin STBSVC solve the singularity problem and improve the aforementioned plane-based clustering algorithms. Experimental results performed on benchmark UCI datasets indicate that the proposed methods outperform other existing plane-based clustering algorithms. Additionally, we also give the application of the proposed method to bio medical image clustering and marketing science. Numerical experiments on real world benchmark datasets show that the proposed models give better generalization perfor mance.