Research On Multiple Kernel Clustering Based On Extreme Learning Machine

Multiple kernel clustering algorithm has better performance than traditional machine learning methods in processing multi-source heterogeneous data.It has attracted the attention of experts in related fields such as data mining and machine learning,and is widely used in medical,commercial,biological,geographic,and Internet,e-commerce,finance,entertainment and other industries.Because multiple kernel clustering takes up a lot of time to achieve a good clustering effect,the existing multiple clustering methods cannot be widely applied in the task of processing multi-source heterogeneous data with high real-time performance.Particularly,it would lead to incomplete kernel matrix when the collected,transmitted or stored data are absent.The existing multiple kernel clustering methods cannot effectively deal with this phenomenon.Therefore,it is of great theoretical significance and value to improve the operation efficiency of multiple kernel clustering and to study the algorithm for processing absent kernel matrix.The main research results are as follows:(1)In order to improve the learning efficiency of multiple kernel clustering,we proposed a two-stage multiple kernel method based on extreme learning machine.It uses several predefined kernel functions of multiple kernel learning methods to capture information in heterogeneous data and learns an optimal representation of heterogeneous data.In the first stage,multiple kernel heterogeneous data is initialized by multiple kernel clustering,and then the extreme learning machine is used to learn the weight coefficient of the base kernel in the second stage.This method combines the multiple kernel clustering with the extreme learning machine.It makes the result of the multiple kernel clustering guide the extreme learning machine to learn the optimal base kernel joint coefficient.In this way,the optimal clustering results can be learned quickly.The experiment results show that the proposed method is about 10 times faster than other multiple kernel clustering algorithms when compared with the existing multiple kernel clustering methods with weak advantages.(2)In order to effectively solve the problem that the multiple kernel clustering algorithm cannot effectively deal with the absence of kernel matrix.An unsupervised multiple kernel method based on kernel completion extreme learning machine is proposed,which can learn unlabeled incomplete data from multiple sources with high efficiency and fast learning speed.The proposed method first initializes the missing values in the kernel matrix with 0,and then uses several base multiple kernel functions to map the data to multiple kernel Hilbert spaces.Finally,iteratively 1)clustering the multiple kernel Hilbert spaces,2)learning the kernel matrix corresponding to the combination coefficients of base kernel function,and 3)inputing the absent values in the kernel matrix.Experiment results show that the proposed method can effectively estimate the missing values in the kernel matrix when the kernel matrix has a low missing rate.When the convergence condition satisfies,the proposed method can output a matrix closed to the original base kernel.At the same time,the cluster analysis result of the data can be obtained.