Tensor Nuclear Norm Based Multi-view Clustering With Adaptive Neighbours Learning

With the rapid development of data acquisition,Internet and sensor technology,the data acquired by people has gradually changed from single view data to multi-view data.Different views can provide useful complementary information for each other.Multi-view learning has become a research hotspot in machine learning,artificial intelligence and other fields.As an important research direction of multi-view learning,multi-view clustering has developed rapidly in recent decades and made great progress.Among them,the adaptive neighborhood multi-view learning algorithm uses iterative method to learn the weight of each view adaptively,and achieves good results.However,like most existing graph based multi-view clustering methods,this method still has the following problems:when calculating different views,it simply adds different views without considering the high-order information and related information between different views.Moreover,this method requires all views to share the same graph,which will lead to over learning.To solve this problem,this paper studies the adaptive neighborhood learning multi view clustering algorithm based on tensor kernel norm(1)In order to solve the problem of over learning and ignoring the high-order information and complementary information between different views,the existing adaptive neighborhood learning clustering methods assume that all views have the same neighborhood,and ignore the high-order information and complementary information between different views.This constraint is relaxed and a third-order tensor is constructed from the graph of different views,and then the adaptive neighborhood learning is carried out by minimizing the tensor weighted kernel norm Based on the high-order information and complementary information between views,two kinds of multi-view clustering algorithms are proposed.Experiments are carried out on ORL,cal101,HW,MSRC and Scene15 multi-view data sets.The experimental results show the effectiveness of the proposed methods.(2)In order to better reflect the differences between different views,a weighted parameter is added to the Laplacian matrix of each graph in adaptive neighborhood learning multi-view clustering.In order to describe the target rank more flexibly,the tensor schatten-p norm is introduced to further improve the performance of multi-view clustering.Experiments are carried out on the above five multi view datasets.Experimental results show that the proposed method improves the performance of multi view clustering.