Semi-Supervised Learning On 1-D Embedding Space

Depending on the development of computers and information technology,people acquire billions of data,such as DNA data,protein sequences,and so on.However not all of the data are clearly labeled.when there are vast amounts of unlabeled data,how to get valuable classification information attracted the attention of researchers.Traditional supervised learning and unsupervised learning could not make full use of labeled and unlabeled data.Semi-supervised machine learning can solve the problem,and it not only makes effectively use of labeled data but also the unlabeled data,to improve the efficiency and accuracy of the classifier,by assisting the training process.We make a study about semi-supervised learning in theory,method and applica-tion,and we focus on a method of semi-supervised learning on 1-D embedding space.The main contribution includes:(1)This thesis introduces the current state of semi-supervised learning.It sees about the appearance and development of semi-supervised learning,and gives a classi-fication and summarization to the existing work.(2)On the problem of semi-supervised classification,we study a ordering scheme based on "shortest possible path".And we propose a novel semi-supervised method of processing the high-dimensional data called 1-D embedding manifold learning,which simply the high-dimensional data processing with maintaining the integrity of the data structure.(3)On the problem of main factors affecting the classification,we mainly think over the assistance of 1-D Multi-Embedding to semi-supervised classification.Tak-ing cubic spline function(CS)and laplacian support vector machine(LapSVM)for interpolation in classification of handwritten digits,we compare the results with traditional SVM classification in the experiment,verifying 1-D embedding method is reasonable and effective.