Research On Clustering Learning Algorithm On Spectral Manifold

In recent years,the data of various forms obtained from multiple data sources show exponential explosive growth,how to effectively extract the inherent properties and rules of data hidden under the cumbersome data representation is a hot topic in the field of ma-chine learning.Because manifold learning can extract meaningful information and low-dimensional representation from the original high-dimensional data,moreover,the assump-tion that the observation data in the multi-manifold model are located or close by multi-ple intrinsic low-dimensional manifolds embedded in the high-dimensional Euclidean space conforms to the learning task of clustering analysis to find multiple disjoint groupings of observation data.Furthermore,learning the manifold structure of data by using spectral analysis is an effective method.Therefore,this paper hopes to realize clustering learning by learning the spectral manifold structure of data.The main work of this paper includes the following three aspects:Firstly,by introducing the multi-manifold structure information of data into tradition-al spectral clustering,we proposed a method named Spectral multi-manifolds embedded clustering(SMEC).The algorithm constructs a similarity matrix which records the location structure information of sample data points in the original observation space and mapping manifolds simultaneously,and further constrains its specification by using the data manifold mapping information,thereby improving the clustering performance by enhancing the judg-ment of the real location of data points.Experimental results on confusing data subsets and real datasets demonstrate the effectiveness of the algorithm.And by further introducing the multi-manifold structure information of data into multi-view clustering,we proposed a method named Self-weighted Multi view Spectral Clustering on Multiple Manifolds(SwMMC).By restricting the similarity matrix,the algorithm can adaptively learn the weight of each view,thus completing the fusion of multi-view data information without parameters,and can directly get the cluster labels of sample data points from this similarity matrix.And the algorithm further improves the learning performance of the algorithm model by introducing the judgment of the manifold structure of each view data into the model as the correction information.Also the algorithm carries out spectral analysis and calculation on the constructed data subset,which reduces the computational burden of the algorithm to a certain extent.Experiments on artificial datasets and real datasets show that the algorithm has better performance than other algorithms.Finally,by introducing deep learning into spectral multi-view clustering,we proposed a method named MultiSpectralNet:Spectral Clustering using Deep Neural Network for Multi-view Data(MvSN).The model provides mapping functions for extracting low-dimensional fusion embedded features and predicting cluster labels for multi-view sample data points.And by using the complementarity of multi-view data information,the method can cor-rect the embedding branch of single-view through the feedback of the data fusion layer to achieve more accurate judgment of the data point position.Besides,the algorithm has the advantages of clustering large-scale datasets and processing out-of-sample-extension.A se-ries of experiments on artificial datasets and real-world datasets have proved the superiority of the algorithm.