Research And Application Of Locality And Sparsity Constrained Adaptive Graph Based Labepropagation

Semi-supervised learning algorithm,as a hotspot of machine learning,takes full advantage labeled samples and unlabeled samples information,so that it has excellent performance in many fields.For semi-supervised learning methods,graphbased semi-supervised learning methods have already turned into one of the most active problems in the arena of machine learning and pattern recognition with the advantages of fast computation speed and high accuracy in recent years.Label propagation,as an effective graph-based semi-supervised classification method,is widely used in image classification,text classification and other tasks.We propose an improved scheme for solving the shortcomings of label propagation algorithm.In addition,we settle the deficiencies of other machine learning algorithms with the model,as follows:First,for graph-based semi-supervised classification method,the construction of graph affects the performance to some extent.Although a large number of graph construction methods have been proposed,the existed methods have the problems of separating graph construction from label propagation process and ignoring the locality and sparsity of data.In order to solve the above problems,we propose a Locality and Sparsity Constrained Adaptive Graph based Label Propagation(LSCAGLP)method.In this method,we combine graph construction with label propagation to form a unified framework.We integrate locality and sparsity of samples into the process of graph construction,making the graph sparser and more discriminant.In addition,an iterative optimization algorithm is proposed to solve the objective function.Experiments show that LSCAGLP algorithm is superior to other comparison methods on four databases.Secondly,ridge regression(RR)algorithm can't use unlabeled samples,while it makes full use of label information of labeled samples.In order to solve the above problems,we combine LSCAGLP with ridge regression algorithm to form a SemiSupervised Ridge Regression with Adaptive Graph-Based Label Propagation(SSRRAGLP)algorithm.In this model,not only does make ridge regression take full use of unlabeled samples information,but also it solves the problem of "out of sample" in label propagation.In addition,an iterative optimization algorithm is proposed to solve the objective function.A large number of experimental results prove the effectiveness of our model.Thirdly,non-negative matrix factorization is an effective dimensionality reduction technology,which discards the label information of samples,so that reduces the ability of feature discriminant.To solve this problem,we present an Adaptive Graph based Semi-Supervised Non-negative Matrix Factorization(AG-SSNMF)algorithm by combining LSCAGLP model with NMF algorithm.It improves performance of the algorithm by introducing the label information of samples.Similarly,an iterative optimization algorithm is used to solve the objective function,and a large number of experiments are carried out to verify the superiority of the algorithm.