Font Size: a A A

Research Of Semi-supervised Classification Learning Framework Based On Multi Assumptions

Posted on:2019-09-13Degree:MasterType:Thesis
Country:ChinaCandidate:Y T ShenFull Text:PDF
GTID:2428330566999352Subject:Software engineering
Abstract/Summary:PDF Full Text Request
Semi-supervised classification(SSC)learning is an important research field in classification learning,which uses the labeled and unlabeled instances simultaneously to alleviate the limitation of instance labels.Manifold regularization(MR)is a classic and powerful framework of SSC learning.However,there are still some issues whith MR: 1)In MR,the labeled instances in semi-supervised are randomly located,they may be in the boundary area or the opposite class,thus label propagation from those labeled instances to their neighbors in MR may mislead the final classification,though the distribution structure of unlabeled samples has been considered.2)The smoothness constraint in MR is implemented over all instance pairs,and actually considers each instance pair as a single operator.However,the smoothness can be pointwise in nature,specially,it shall inherently occur “everywhere” to relate the behavior of each instance point to that of its close neighbors.As a result,the main work in this paper will focus on the following two parts:Firstly,to reduce the impact of extremely scarced or misleading labeled instances by expanding the labeled instance set,a novel label-expanded MR framework(Label-expanded Manifold Regularization for semi-supervised classification,LE_MR for short)is developed.In LE_MR,a clustering strategy such as KFCM is first adopted to discover the high-confidence instances,i.e.,instances in the central region of clusters.Then those instances along with the labels are adopted to expand the labeled instances set,which alleviates the lack of labeled instances and improves classification performance of MR.Experimental results on real data sets show that LE_MR based on extended labeled samples can effectively enhance the learning performance of MR.Secondly,by considering the smoothness of individual instances rather than the instance pairs,a novel pointwise MR framework(Pointwise Manifold Regularization for semi-supervised learning,PR_MR for short)is developed,in order to preserve the pointwise nature of smoothness and reduce the misleading of the instance pairs.In PR_MR,the pointwise nature of smoothness is preserved by considering individual instances rather than the instance pairs,at the same time,the importance of individual instances described by local density can be introduced for improving the classification performance of MR.Experimental results on real data sets show that pointwise MR can help improve the learning performance of the MR framework.
Keywords/Search Tags:Semi-supervised Classification, Manifold Regularization, membership degree, pointwise smoothness, local density
PDF Full Text Request
Related items