Research On Multiple Kernel Clustering Algorithm Based On Local Learning

Data mining is a subject to mine valuable information from a large number of data or databases,which has been applied in many fields.Clustering analysis,as an irreplaceable mining technology in data mining,is also widely used.Clustering analysis divides samples into different clusters or subsets according to similarity,which makes the samples in different clusters have great differences.In recent years,kernel method has been widely used in clustering tasks because of its advantages in nonlinear pattern analysis tasks.However,the performance of kernel clustering largely depends on the setting of kernel function and parameters.Therefore,the direction of multiple kernel clustering has emerged.In recent years,the multiple kernel clustering method has also been widely concerned by researchers,and continuous progress has been made in this field.This paper is based on the research of multiple kernel clustering algorithm.We find that most of the existing multiple kernel clustering algorithms do not fully consider the local manifold structure,but the characterization of the local manifold structure is very important to improve the clustering performance.To address this problem,we proposes two multiple kernel clustering algorithms based on local learning,and further develops a multiple kernel clustering algorithm analysis system.The specific work content is described as follows:(1)A manifold adaptive multiple kernel K-means algorithm(MAMKKC)is proposed.The multiple kernel clustering algorithm based on K-means aims to improve the performance of kernel K-means clustering by integrating a set of cores.However,most of the existing multiple kernel k-means algorithms make use of the nonlinear relationship in the kernel,and the local manifold structure in the multiple kernel space is not fully considered.In this paper,manifold adaptive kernel is used instead of the original kernel to integrate the local manifold structure of the kernel.Therefore,the induced multiple manifold adaptive kernel not only reflects the nonlinear relationship,but also reflects the local manifold structure.In addition,an iterative algorithm is proposed to solve the optimal solution.The experimental results show that this method is superior to many advanced algorithms on multiple benchmark datasets.(2)An optimal neighborhood based multiple kernel concept decomposition algorithm(ONMKCF)is proposed.This algorithm is a new multiple kernel concept decomposition algorithm for data representation and clustering.Firstly,in the multiple kernel framework,the kernel concept decomposition is extended to alleviate the problem of kernel selection.Then,considering that the nonlinear relationship in kernel space does not capture the manifold,we learn the optimal neighborhood graph from multiple candidate graphs induced by multiple kernels to reflect the internal structure of the manifold.At the same time,a new rank constraint is imposed on the Laplacian matrix of the multiple candidate graph matrix,so that the connected components in the multiple candidate graph matrix are exactly equal to the cluster number.In addition,an iterative algorithm is proposed to solve the optimal solution.The experimental results show that this method is superior to many advanced algorithms on multiple benchmark datasets.(3)This paper designs and develops a multiple kernel clustering algorithm analysis system,which includes datasets selection,multiple kernel clustering algorithm selection and algorithm results display.This system integrates the two algorithms and comparison algorithm proposed in this paper.The system can select different datasets for experiments,and has good usability.To sum up,this paper proposes two multiple kernel clustering algorithms based on local learning.Clustering analysis plays an irreplaceable role in both theory and practical application,so the two clustering algorithms proposed in this paper are of certain significance and value.