| In the era of big data,data usually comes from different fields or observed from different angles.For example,the images are described by different types of features;the same news is reported in different languages;one object can be captured by multiple cameras located in different angles at the same time,which are called the multi-view data.Compared with the traditional single-view data,the multi-view data has the more comprehensive information.In recent years,the cluster analysis of multi-view data has gradually become a research topic in the field of the machine learning and the data mining,and is widely used in the fields of the face recognition,abnormal behavior recognition and crossmedia retrieval.K-means clustering algorithm,as a basic unsupervised learning method,has been widely studied due to its high efficiency and effectiveness.However,the Kmeans algorithm divides clusters by measuring the distance between samples,which is difficult to deal with the high-dimensional data.For the complex multi-view data,it is more difficult to deal with its non-clustered data distribution.Aiming at the problems and defects of the traditional K-means clustering methods,we studies the deep K-means clustering algorithm of high-dimensional data and the multi-view clustering method that integrates the complementarity and consistency of different views,and achieves the following innovative results:Firstly,a real deep K-means with multiple auto-encoders is proposed.Aiming at the problem that the traditional K-means algorithm is difficult to deal with the highdimensional data,using the powerful nonlinear representation ability of neural network,the centroid of each cluster is acted by one auto-encoder rather than the constant vector,and the clustering is achieved by assigning sample points to the auto-encoder that can best reconstruct.Secondly,a adaptive K-Multiple-Means for multi-view clustering is proposed.Aiming at the problem that the traditional multi-view K-means clustering algorithm can-not deal with the non-clustered dataset,it is proposed to use the multiple sub-cluster centers to capture the distribution of each cluster in each view.In addition,a new multiview fusion weighting strategy is adopted to automatically assign the optimal weight to each view,so as to effectively fuse the complementary information between different views and obtain a unified representation of multi-view data.Thirdly,a multi-view subspace clustering network with second-order graph information is proposed.The existing multi-view K-means algorithm is difficult to describe the complex relationship of the high-dimensional multi-view data.In order to this,we proposed to introduce the graph structure information between the data into the deep multi-view subspace network,which combining the first-order and second-order graph information of the multi-view data into the self-expression layer of network to guide the learning of the subspace representation.In addition,the discriminative constraints are imposed on different views,so as to obtain the better clustering performance.In order to verify the effectiveness of the proposed method,the massive experiments are conducted on several public multi-view datasets.The experimental results show that the clustering algorithm proposed can achieve good performance. |
