Research On Multi-view Clustering Algorithms Based On Graph Learning

Clustering is a hot topic in machine learning,many algorithms have achieved good performance,but it is limited to a single view,with the explosive growth of data,the performance of the former single-view clustering model to deal with big data is obviously reduced,and more and more scholars begin to study the multi-view clustering problem.For the multi-view clustering problem,it is only simple to join each view at first,and this method ignores the complementarity between the views.After that,some methods establish a unified function and optimize the multi-view data to improve the generalization performance.In recent years,most methods focus on mining the consistency between different views,neglecting the unique attributes of different views,and the constructed graphs are not optimal graphs,and some algorithms have high computational complexity.In view of the above problems,the main work of this paper is as follows:（1）A multi-view clustering algorithm based on diversity consensus spectral embedding learning is proposed.In order to learn the consensus graph better,the algorithm automatically learns the consensus spectral embedding matrix and the discrete clustering label matrix under the consideration of the diversity of views.By comparing with other algorithms on real data sets,the superiority of the propose algorithm in improving clustering performance is proved.（2）We propose a novel robust consistent graph learning（RCGL）method.More specifically,RCGL combines multi-view inconsistency and matrix factorization into the fusion graph learning,and learns the consensus similarity graph dynamically.By introducing the orthogonal and nonnegative constraint,the interpretable and robust clustering results can be obtained without any postprocessing steps.In addition,an alternative optimization strategy is presented to optimize the objective function.Experimental results on the real data sets demonstrate that the proposed method achieves the comparable or even better clustering performance than the state-of-the-art multi-view clustering methods.（3）A clustering algorithm based on sparse consensus graph decomposition is proposed.The algorithm is tested on single-view data,and then extended from single-view data to multi-view data.The algorithm takes into account the different contribution of different views to the final result,gives each view appropriate weight,at the same time,makes use of the L_{2 1,}norm to obtain the consensus graph with better performance,learns the non-negative representation matrix on the basis of the consensus graph,and reveals the cluster result directly after alternation iteration.In addition,an update iterative algorithm is proposed and tested on a large number of data sets to verify the effectiveness of the algorithm.