Research On Spectral Learning Algorithm For Multiple Source Data

In the process of human cognition and transformation of the world,it is often necessary to use multi-source data.Compared with single source data,multi-source data contains more abundant information and knowledge.Through the research and analysis of multi-source data,a more comprehensive and objective cognition of things can be obtained.The machine learning algorithm for multi-source data can improve the learning ability of the algorithm by utilizing the hidden knowledge in the multi-source data.However,traditional machine learning methods are generally oriented to single-source data modeling,and multi-source data makes traditional machine learning methods face new challenges.This dissertation studies the spectral learning algorithm for multi-source data,and the main work of this paper includes:On the basis of the complementarity and consistency of multi-source data,a new model of multi-source data fusion learning is established.The new model can represent the internal structure of multi-source data more comprehensively by combining global spectral embed-ding fusion and local spectral embedding fusion.For the new model,this paper presents the corresponding optimization algorithm.Experiments show that the proposed algorithm is effective and can improve the performance of isomorphic complete multi-source data fusion clustering.In order to reduce the complexity of fusion learning model and improve the learning ef-ficiency,a spectral density representation method is proposed.Firstly,the consistent spectral density representation is obtained,and the optimal density representation was obtained by linear combination of spectral density representations of each data source.The density peak clustering algorithm is used to complete the multi-source data clustering task.Experiments on multi-source datasets show that the algorithm has the ability of multi-source learning and high efficiency,and the ability of processing noise data.In order to solve the problem of missing data,a spectral recovery method is proposed.Firstly,the data sources are filled by establishing projections between different data sources.Then the spectral properties of the operator are used to recover the missing data.Thus,the true and complete adjacency matrix of each data source is obtained.Based on the real and complete adjacency matrix of each data source,the fusion model of multi-source data is established.Experiments on incomplete multi-source data sets show that the algorithm can solve the problem of missing data in multi-source data and realize the fusion learning of incomplete multi-source data.In order to solve the problem of missing association in heterogeneous multi-source data,it is proposed to construct the association between data sources by using the sample corre-spondence and feature correspondence.The association relation in traditional multi-source learning is mainly reflected in the sample correspondence.However,when the sample cor-respondence is missing,the traditional methods cannot establish the learning model.Based on the method of manifold alignment,this paper combines two kinds of correspondence to realize the fusion learning of heterogeneous multi-source data.Experiments on heteroge-neous multi-source datasets show that the algorithm performs better and more stably than traditional multi-source learning methods when the sample correspondence is missing.The innovation of this paper mainly includes:1.Aiming at the problem of multi-source data fusion learning,this paper proposes the multi-source spectral embedding fusion learning algorithm.2.Aiming at the problem of multi-source data representation,this paper proposes the multi-source spectral density representation learning algorithm and the multi-source spectral recovery learning algorithm.3.Aiming at the problem of sample correspondence missing in multi-source data,this paper proposes the multi-source manifold alignment spectral learning algorithm.