Local learning integrating global structure for large scale semi-supervised classification

In recent years, semi-supervised learning algorithms have aroused considerable interests from machine learning fields because unlabeled samples are often readily available and labeled ones are expensive to obtain. Graph-based semi-supervised learning has been one of the most active research areas. However, how to speed up these methods for handling large scale datasets is still a challenge. In this paper, we apply the clustering feature tree to large scale graph-based semi-supervised learning and propose a local learning integrating global structure algorithm. By organizing the unlabeled samples with a clustering feature tree, it allows us to decompose the unlabeled samples to a series of clusters (sub-trees) and learn them locally. In each training process on sub-trees, the clustering centers are chosen as frame points to keep the global structure of input samples, and propagate their labels to unlabeled data. We compare our method with several existing large scale algorithms on real-world datasets. The experiments show the scalability and accuracy improvement of our proposed approach. It can also handle millions of samples efficiently.