Semi-supervised learning via complex networks: network construction, selection and propagation of labels and applications

Abstract

Graph-based semi-supervised learning (SSL) is a promising paradigm for modeling the manifold in multidimensional data space, and is effective in propagating a small number of initial labels to a large amount of unlabelled data. This approach has been used in a variety of applications, such as image segmentation and annotation, human speech recognition, text classification, etc. Recently, some authors observed the importance of the generated network for the label propagation process, but other aspects were still little investigated, for example, the selection of the initial labeled samples or the topological characteristics of the network. Thus, the objective of this project is to investigate in depth all the steps involved in SSL, including the selection of initial labels, the selection of the network to be built and the method for propagating labels, and propose new approaches to improve this process. Challenges of the area will also be addressed such as contaminated or unbalanced labels, large databases and applications in text mining, climate data analysis and data augmentation for images. (AU)