A Study On Some Problems Of Semi-supervised Learning

Due to limited resources, obtaining vast labeled examples is difficult for practical applications. With a few labeled examples and large number of unlabeled examples, how to improve the performance of supervised and unsupervised learning is a key problem of semi-supervised learning. Studies on semi-supervised learning are valuable. On the one hand, semi-supervised learning uses vast unlabeled examples to assist in training a more effective classification. On the other hand, the results of unsupervised learning are more effective and reasonable with the guidance of prior knowledges (e.g., labeled examples and pairwise constraints).We discuss the studies on clustering of complicated dataset, classification and outlying trajectory detection in the area of semi-supervised learning. Three novel algorithms are proposed by analyzing the deficiency of typical algorithms:(1) The algorithm of Semi-supervised Clustering for Dataset with Complicated Structure (SCDCS), which can find the clusters of arbitrary shapes, sizes, and densities, even in the presence of noise and outliers.(2) The algorithm of A Novel Semi-supervised Classification with Supervised Clustering (N2SC), which can reduce interference of mislabeled data and enhance the performance of classifier.(3) The algorithm of Semi-supervised Trajectory Outlier Detection (Semi-TOD), which combines the prior knowledge of a few outlying trajectories and multi-view outlier detection to detect available outlying trajectories as well as reduces the negative impact of practical applications and human factors.The three novel algorithms proposed were tested on some public datasets and real datasets. The Experimental results show that our studies on semi-supervised learning are feasible, which will extend the existing studies and have certain application value.