Researches On Semi-supervised Learning

information processing. In this thesis, we study on an important branch of machine learning called semi-supervised learning, which mainly focuses on the problem of learning with both labeled data and unlabeled data. We propose some systemic researches about semi-supervised learning in theory, method and application. More concretely, the main contribution includes:1. This thesis extensively sees about the current state and future trend of semi-supervised learning. It reviews the appearance and development process of semi-supervised learning, and gives a summarization to the existing work. We primarily analyze current researches about semi-supervised learning in theory, method and application and conclude the shortages of some methods.2. On the problem of semi-supervised dimensionality reduction, we concern the methods based on feature selection. And we propose a novel semi-supervised method of feature selection, which is called Semi-supervised Discriminant Feature Selection. Experiment shows the effectiveness of SDFS.3. On the problem of expansion of semi-supervised learning, we mainly think over the assistance of a new kind of data called the Universum. Combining the advantages of traditional semi-supervised methods and the Universum and basing on linear regression and graph optimization framework, we present a novel method for classification called Semi-Supervised Classification with the Universum. The effectiveness is verified by both simulation and real-world data.