Graph Based Semi-supervised Learning And Applications

Supervised learning and unsupervised learning are two mostly used learning method in machine learning, however when there are only limited labeled samples, it is hard for supervised learning to learning the real distribution of the whole dataset. Unsupervised learning does not need the labeled samples, but for the dataset with complex spatial distribution, unsupervised learning fails to obtain a good result too. To solve these problems, semi-supervised learning has been paid closed attention in recent years in the area of machine learning. Semi-supervised learning combines the advantages of supervised learning and unsupervised learning, and use the large amount of the unlabeled samples to improve the performance of the algorithm.This paper has remote sensing image classification and segmentation as the application background. The traditional graph-based semi-supervised learning method and one classical particle swarm optimization (PSO) classification method as well as a semi-supervised clustering method—semi-supervised spectral clustering was studied. The improved methods have been proposed and were successly used in hyperspectral image classification and SAR image segmentation. The main contributions can be summarized as follows:(1) A new method of discriminative graph-based semi-supervised classification method is proposed. The discriminative information was introduced when constructing the graph, making full use of the limitied labeled samples to construct a discriminative graph. And then the traditional graph-based semi-supervised learning framework was used on the proposed discriminative graph for classification, and the classification result was improved compared some other sem-supervised method.(2) A semi-supervised clustering method-l1-graph based semi-supervised spectral clustering was proposed. The nystrom method was used to accelerate the construction of the l1-graph. The l1-graph was constructed based on the sparse representation, and has been proved a good performance on spectral clustering. We used the pairwise constraints on the l1-graph and proposed the l1-graph based semi-supervised spectral clustering. However when there are lots of samples, the computation cost of constructing the l1-graph is high, so we proposed to use the nystrom method to construct the graph. The proposed method was used on SAR image segmentation and obtained a good segmentation results.(3) A semi-supervised PSO classification method was proposed. For the traditional PSO classification method, it is difficult to find the optimal positions of the particles when there are only limited labeled samples. The proposed semi-supervised can use large amount of unlabeled samples to search the optimal particles positions in the solution space, and finally improve the classification results.