LNP-Based Semi-supervised Learning Algorithms

Semi-supervised learning (SSL), which uses limited labeled samples with assistance of a great many of unlabeled samples, could effectively improve learning accuracy. Though using the supervised information, SSL learns the whole samples or merely the unlabeled sample subset. SSL not only reduces the workload and the error from incorrect manual labeling, but also remarkably improve the accuracy. Therefore, the research on SSL is of great importance both in theory and practice.Firstly, a new incremental classification algorithm based on LNP is proposed:Incremental LNP. It takes advantage of the fact that there does exist a hidden data selection principle in multi-class LNP, and is implemented in the style of incremental learning. The experimental results demonstrate that the new algorithm not only has a good ability of convergence, but also obviously improves the accuracy when compared with that of LNP; in addition, a study of the two parameters involved in the new algorithm is made, which leads to the conclusion that when proper parameters are selected, higher accuracy will be obtained within finite iterations.Secondly, the Incremental LNP algorithm is applied to deal with semi-supervised dimensionality reduction; thus, the obtained dimensionality reduction algorithm is on the process of iteration, also we can get the result after dimensionality reduction at any arbitrary iteration. In the experiments, the Nearest Neighborhood classifier is used to categorize the lower-dimensional data, and the result turns out to be well convergent.Finally, a semi-supervised learning model based on separability criterion and LNP is put forward: the parameter is chosen by separability criterion firstly; and then LDA is employed to reduce the dimensionality of the samples labeled by LNP; finally, the Nearest Neighborhood classifier is used to test the performance of the model. This model makes the best of both LNP and LDA; also it is capable of handling the problem of"a single training sample"which LDA fails to; what's more, it contains three separability criteria to select the parameter of LNP. In the experiments, though comparing our methods with some classical ones, the experimental results validate the effectiveness and superiority of the proposed method; meanwhile, the semi-supervised learning models based on three separability criteria are also compared with each other and analyzed in the thesis.