Semi-Supervised Classification Method Based on Spectral Clustering

摘要: With the rapid development of data collection and storage technology, there are plentiful unlabeled but very few often expensive labeled in real-word applications. Thus, semi-supervised learning algorithms have attracted much attention. In this paper, we propose a new classification algorithm benefiting from spectral clustering called SC-SSL. First, introduce to partition all into clusters. Second, build classifier using predict probabilities (weights) classes that each instance belongs for cluster. Third, cluster, add those instances whose labels with maximum weights as same cluster label data. Fourth, terms set, reconstruct classifier. We repeat above processing steps 2 3 till meeting stopping condition. Finally, extensive experiments reveal our SC-SSL can sufficiently use information get robust by clustering, it maintains higher accuracy compared several well known algorithms.