A simple graph-based semi-supervised learning approach for imbalanced classification

Abstract

Graph-based Semi-Supervised Learning (GSSL) methods aim to classify unlabeled data by learning the graph structure and labeled data jointly. In this work, we propose a simple GSSL approach, which can deal with various degrees of class imbalance in given datasets. The key idea is to estimate the class proportion of input data in order to enhance the discriminative power of learned smooth classification function on the graph. Moreover, it has interesting connections to the regularization framework, the Markov stability for graph partition and the group inverse of normalized Laplacain matrix. For classification problems, experimental results demonstrate our approach can achieve promising performance on several datasets with varying class imbalance.