Research On Semi-supervised Dimension Reduction Methods Based On Structured Graph Learning

With the rapid development of information technology,high-dimensional features can provide more comprehensive descriptions for the objective world.However,they also bring considerable computing burden and high storage cost.Therefore,the dimension reduction methods that project the original data features into the low-dimensional space with original information preserving have attracted extensive attention of researchers.According to the label usage,dimension reduction approaches can be roughly divided into supervised,unsupervised,and semisupervised.Among them,semi-supervised dimension reduction comprehensively adopts labeled data and unlabeled data to achieve dimension reduction process,which has the advantages of both supervised and unsupervised methods.As an effective semi-supervised dimension reduction method,graph-based semi-supervised dimension reduction aims to preserve the structural information of all data points by the form of similarity graph,and then improve the feature projection performance through label propagation on the similarity graph.In this method,label propagation is performed to transfer the discriminative semantics from the labeled samples to the unlabeled samples through the graph.Therefore,the performance of graph-based semi-supervised dimension reduction approach heavily depends on the quality of the graph.The existing graph-based dimension reduction methods mainly suffer from three important problems:(1)Graph construction and dimension reduction are separated into two independent steps.Moreover,real-world data usually contain a lot of redundant and noise information,which will reduce the quality of the pre-constructed graph.Under such circumstance,the label propagation on the low-quality graph may generate unreliable labels for unlabeled samples.(2)Although several approaches have achieved certain success by integrating these two independent steps into a joint learning framework,the learned graph with sub-optimal structure still has limited capability to support the label propagation process.Therefore,unreliable labels will be generated and thus affect the feature projection learning process.(3)Most of the existing methods only consider the single layer semantic propagation on the graph.Besides,they generally adopt a mandatory semantic consistent constraint that forces the learned labels of labeled samples to be directly equal to the given labels.However,the given labels are discrete and the learned pseudo labels are continuous.Directly forcing them to be equal will cause semantic loss and thus reduce the subsequent projection performance.To solve these problems,two semi-supervised dimension reduction methods are proposed:(1)This thesis proposes an effective label propagation with structured graph learning method for semi-supervised dimension reduction.In this model,label propagation,semi-supervised structured graph learning and dimension reduction are simultaneously performed in a unified learning framework.Different from the previous methods,a semi-supervised structured graph learning method to characterize the intrinsic semantic relations of samples more accurately is proposed.Further,the method assigns different importance scores for the given and learned labeled samples to differentiate their effects on learning the feature projection matrix.The semantic information can be propagated more effectively from labeled samples to the unlabeled samples on the learned structured graph.And a more discriminative feature projection matrix can be learned to perform the dimension reduction.(2)A semi-supervised dimension reduction model based on a dual-layer semantic propagation is proposed.This model simultaneously performs semantic propagation on a structured graph and considers the soft semantic alignment to obtain a more discriminative projection matrix.The dual-layer semantic propagation strategy can achieve effective label propagation.In addition,using the soft semantic alignment to replace the traditional mandatory semantic consistent constraint can effectively reduce semantic loss during the label propagation and accelerate the optimization process.Experiments demonstrate the proposed two semi-supervised dimension reduction methods based on structured graph learning can achieve state-of-the-art performance on both accuracy and learning efficiency.