Manifold Correlation Graph for Semi-Supervised Learning

Abstract

Due to the growing availability of unlabeled data and the difficulties in obtaining labeled data, the use of semi-supervised learning approaches becomes even more promising. The capacity of taking into account the dataset structure is of crucial relevance for effectively considering the unlabeled data. In this paper, a novel classifier is proposed through a manifold learning approach. The graph is constructed based on a new hybrid similarity measure which encodes both supervised and unsupervised information. Next, strongly connected components are computed and used to analyze the dataset manifold. The classification is performed through a voting scheme based on primary (labeled) and secondary (unlabeled) voters. An experimental evaluation is conducted, considering various datasets, diverse situations of training/test dataset sizes and comparison with baselines. The proposed method achieved positive results in most of situations.