Nowadays,the rapid development of information technology has resulted in a variety of data,and people use these data to process all kinds of information.Different forms of description information for the same object are called multi-view data.Clustering is a method to find a certain internal law from these massive data.Multi-view clustering is to integrate the information of multi-views in a certain way,and give full play to the information advantages of each view to perform clustering,which is deeply loved by scholars.However,multi-view clustering still has some shortcomings,such as data noise or feature redundancy,and lack of robustness.In response to these shortcomings,this article has conducted in-depth explorations on how to extract features that are easier to cluster and how to enhance the robustness of clustering.A multi-view clustering algorithm based on subspace fusion and robust deep multi-view subspace fusion are proposed.Clustering Algorithm.details as follows:(1)This paper proposes a multi-view clustering algorithm based on subspace fusion.For many computer vision problems,the feature representation of data is usually distributed in some low-dimensional subspace.Subspace clustering is a method to find such low-dimensional subspaces and correctly cluster data points.Multi-view subspace clustering is a current research hotspot.How to fully integrate the complementary information of multi-views is a major difficulty.The multi-view subspace clustering method proposed in this paper discards some high-dimensional redundant information,performs deeper denoising on the data,and directly fuses the subspace containing the distribution information of data clusters to achieve more accurate Extract consistent distribution information from multiple views.(2)This paper proposes a robust multi-view depth subspace fusion clustering algorithm.Deep learning has already shined in the field of clustering.In this paper,the multi-view clustering algorithm based on subspace fusion is extended to the depth aspect.In this paper,the original data of each view is passed through an autoencoder to obtain the potential representation of its depth,and then the potential representation of each view is merged into a common potential representation,the subspace representation matrix is found,and the common Subspace means to do spectral clustering.Since some data collected in real life may be damaged,in order to cope with this problem and enhance the robustness of the model,this article adds a bias term to the reconstruction process of the original data of each view and the self-expression process of the latent representation.The model is guaranteed by double robustness.And this paper uses a smoothing technique to eliminate the non-differentiability of the norm,and smoothly approximate the norm to generate a differentiable network,so that back propagation can be used to update the network parameters.The proposed method has been experimented on three widely used multi-view benchmark datasets to verify the robustness and effectiveness of the proposed model. |