Semi-supervised Learning Using Local Regularizer and Unit Circle Class Label Representation

Abstract

Semi-supervised learning, which aims to learn from partially labeled data and mostly unlabeled data, has been attracted more and more attention in machine learning and pattern recognition. A novel semi-supervised classification approach is proposed, which can not only handle semi-supervised binary classification problem but also deal with semi-supervised multi-class classification problem. The approach is based on local regularizer and unit circle class label representation. The former is minimized so as to cause the class labels to have the desired properties. The latter utilizes two-dimensional vector evenly distributed in circumference of unit circle to represent class label, so multi-class classification can be performed only once. Comparative classification experiments on some benchmark datasets validate the effectiveness of the presented approach.