A novel semisupervised classification method via membership and polyhedral conic functions

Abstract

In real-world problems, finding sufficient labeled data for defining classification rules is very difficult. This paper suggests a new semisupervised multiclass classification method. In the initialization, new membership functions are defined by utilizing the labeled data's medoids and means. Then the unlabeled points are labeled with the class of the highest membership value. In the supervised learning phase, separation via the polyhedral conic functions (PCFs) approach is improved by using defined membership values in the linear programming problem. The suggested algorithm is tested on real-world datasets and compared with the state-of-the-art semisupervised methods. The results obtained indicate that the suggested algorithm is effective in classification and is worth studying.