In recent years,the theories and applications of machine learning are quickly developing and consistently improving people’s daily life.At present,machine learning algorithms have been widely used in the fields of people’s livelihood,healthcare,security,and national defense,such as face recognition-based access control and security systems,fingerprint recognition-based payment systems,image retrieval,and Google driverless car.With the rapid development of computer hardware,network and data acquisition equipment,it brings great convenience to people in obtaining abundant data resources.A large amount of data contains rich information,which helps the machine to learn more reasonable models in principle However,such large amount of data also brings problems such as data complexity increasing,information overwhelming and data redundancy.Certain data may contain a lot of noise due to irregular collection methods or abnormal collection environment.Thus,it may be difficult to learn a robust model because of the redundant and noise features.How to extract the most valuable information from different tasks for further analysis has become the focus and difficulty in the field of machine learning and pattern recognition.Feature extraction is the process of extracting useful information from data.An effective feature extraction method not only greatly reduces the dimension of data and thus improves the efficiency of algorithms,but also helps to learn more robust models for different tasks and thus improve the generalization capability of the models.Therefore,the research of feature extraction is of great significance.At present,graph constraint-based models are widely used in pattern extraction and data analysis.Nevertheless,the complexity of data brings new challenges to analysis because data processing may come from single or multiple views.Although these methods have achieved good results in data analysis,they have some disadvantages:(1)most dimension reduction methods only use a single projection matrix to map high-dimensional data into low-dimensional subspace.Projection needs to learn low-dimensional subspace while preserving important features and internal structure of data.(2)Most methods separate graph learning and pattern extraction as two different steps,i.e.,using the local geometry information of data to construct graph,and then using projection learning to extract data patterns.However,such two separate steps are difficult to ensure that the constructed graph is suitable for data pattern extraction,thus it is difficult to guarantee the overall optimization of algorithms.(3)Most methods only consider single-view graph learning.However,as the size and type of data increase,the number of various graphs is increarsing.The existing methods seldom use multi-view learning to construct a correct graph,and then accurately characterize data structure for effective data analysis.(4)Most methods do not eliminate the influence of data noise in the process of data analysis causing less robustness issue.This paper takes a graph constraint-based model as the learning object,uses graph embedding technology,and proposes a series of new models to improve the learning efficiency and robustness of the graph constraint-based model.At the same time,we extend single-view graph embedding technology to multi-view environment and propose a multi-view graph embedding technology to expand the scope of data processing.Specifically,there are a number of contributions as follows:Firstly,this paper proposes a robust analysis framework for adaptive locality preserving,which simultaneously performs graph learning and data feature extraction.Thus it can ensure the overall optimization of algorithms.Specifically,the framework has the following advantages:First,the framework uses a sparse matrix to fit the noise information of data,and then improves the robustness of models.Secondly,the framework guides projection learning by using local structure information and label information of data,while adaptively learning a local structure map to constrain models to avoid overfitting.Finally,the framework learns another projection matrix to maintain the discriminative information of data.In addition,this method enables models to adaptively select most important features in the feature extraction process by imposing L2,1 norm constraints on the projection matrix.A plenty of experiments show that this method can extract more discriminative features and can effectively improve classification accuracy.Secondly,a new unsupervised dimension reduction method,as a relaxed sparse locality preserving projection method,is proposed.This method uses two projection matrices rather than a single projection matrix to reduce the computations on single matrix when processing data,which makes the two matrices have a greater degree of freedom to better maintain the sparsity and local structure of the projected data.This reduces the loss of information of samples during dimension reduction.The similarity matrix of two projection matrices is learned by applying sparse representation,and the local structure is retained in a linear manner.Therefore,the two projection matrices have a similar structure,i.e.,the local manifold structure of the data.In order to solve the difficulty of algorithm optimization,we also propose an efficient iterative algorithm with fast convergence.Experimental results on six datasets demonstrate the effectiveness of the proposed method.Finally,this paper proposes a novel multi-view graph learning method based on diversity-promotion,which further expands single-view graph learning to multi-view graph learning.Specifically,the proposed method constructs a graph for each view and uses adaptive weight linear approximation techniques to enable view graph to adaptively approximate final unified graph so that the final learned graph does not deviate from each view graph.In this method,graph learning is incorporated into a data label transfer model.After that,a generalized framework of joint multi-view graph learning and label transfer is constructed.In order to effectively reduce information redundancy,the framework further considers the difference of each view graph,distinguishes similar view graphs by using an adaptive weight coefficient,and assigns greater weights to larger view graphs,to ensure that final learned graph can accurately depict the intrinsic geometry of data.A number of experiments show that this method not only learns an accurate graph,but also accurately transfers data labels. |