Research On Multi-view Kernel K-means Algorithm Using Weighted Entropy

In recent years, the diversity of data makes traditional clustering algorithms have been unable to meet the requirements of data analysis, so multi-view clustering algorithm was proposed. Existing multi-view clustering technology is mainly divided into three categories, namely the co-training algorithm, the clustering algorithm based on multi-core and the multi-view clustering algorithm based on subspace. However, these algorithms mainly focus on multi-feature dataset, and research about the effect of unrelated or noise perspective on the clustering results are not sufficient. In order to solve this problem, related research is done in this paper.Firstly, the multi-view algorithm convergence was studied. Multi-view algorithms can substantially improve the classification and clustering results, But only these algorithms are convergent, can ensure the effectiveness. In this paper, an analysis of the convergence performance is conducted for a class of multi-view kernel k-means(MVKKM) utilizing the Zangwill convergence theorem. It shows that under certain conditions the iterative sequence generated by a MVKKM converges, at least along a subsequence to either a local minimal or a saddle point of the objective of the algorithm.Secondly, the algorithm kernel-based weighted multi-view clustering is improved.In multi-view clustering based on view weighting, weight value of each view products great influence on clustering accuracy. Aiming at this problem, a multi-view clustering algorithm named Entropy Weighting Multi-view Kernel k-means(EWKKM) is proposed, which assigned a reasonable weigth to each view so as to reduce the influence of noisy or irrelevant views, and then to improve clustering accuracy. In EWKKM, different views were firstly represented by kernel matrix and each view was assigned with one weight. The, the weight of each view was calculated from the corresponding information entropy. Finally the weight of each view was optimized according to the defined optimized objective function, the multi-view clustering was conducted by using the kernel k-means method.Finally, based on artificial data sets and real data sets,the experiments were conducted, experimental results verified the effectiveness of the algorithm.