Research On Multi-View Incremental Clustering Based On Multiple Kernel Learning

People's life is filled with a variety of data,of which a majority of data contains very important information.Data mining is a technique to extract the important information from massive data.As a useful method in data mining,clustering analysis can effectively divide the data into different clusters according to the similarity,which can easily find out the inherent distribution of data objects.The data objects in each cluster have similar information,and there is different information among the data objects of different clusters.With the data being more complex,the traditional clustering algorithm cannot analysis data sets from multiple angles,and then multi-view clustering algorithms emerge as the times require and become research hotspot.A multi-view data set is composed of data that describes one thing from different angles.The multi-view clustering methods cluster the objects by combining the useful information of all views,and then obtain the final clustering result.Kernel function is an effective method for handling the linear non-inseparable data,and MKL is an improvement on kernel function,using a combination kernel function to replace the single kernel function,which is a linear combination of base kernels.And by adjusting the weight of each kernel,making it suitable for different types of data sets.Based on the idea of MKL,this thesis proposes a multi-view clustering based on variable weight and MKL.This algorithm introduces variable weights to measure the contributions of different views for one instance,and the product of weights of each instance is equal to one.So the clustering quality would be improved by increasing the weight of the view with high contribution.Meanwhile,an improved weighted Gaussian kernel function is introduced to solve the problem of kernel function selection.Finally,the weights and cluster results are obtained by iterative method.Experiments on multiple data sets show that this algorithm has higher clustering quality.The incremental clustering algorithm is a classical method of dealing with dynamically increasing data.It can cluster the new data one by one or by batch on the basis of the existing clustering results,so as to avoid a mass of redundant computation and improve the clustering efficiency.In this thesis,the incremental clustering is applied to multi-view data,and a multi-view incremental clustering algorithm based on kernel K-means is proposed.This algorithm divides the data set into a plurality of data chunks.The multi-view kernel K-means clustering method is used to cluster each data chunk,and using a set of cluster centers to reprensent each chunk.Finally,the cluster centers of all chunks are marged,then the final clustering results are obtained by clustering again.Experiments on several data sets show that the proposed algorithm not only guarantees the clustering quality but also reduces the clustering time.