Large-scale multi-view subspace clustering via embedding space and partition matrix

Abstract

Multi-view subspace clustering (MVSC) has attracted increasing attention because it can extract information from multiple views and explore the underlying structure. In general, most of the existing anchor strategies solve the problem of excessive complexity, but there is a loss of information in the process of affinity graph passing to spectral clustering. In addition, the noise in the original data leads to the learned anchor graph not representing the data features adequately. To solve the above problems, this paper proposes a Large-Scale Multi-View Subspace Clustering via Embedding Space and Partition Matrix(LMVSC-EPM) algorithm that preserves the distribution of the original data by embedding matrix mapping. The algorithm utilizes the centroid matrix and the clustering assignment matrix to derive the clustering results directly. Specifically, LMVSC-EPM employs embedding matrices to map the raw data into the embedding space and adaptively learns anchors using a view-sharing anchor strategy. Moreover, the non-negative orthogonal matrix is adopted to assign the results to the clustering assignment matrix, which avoids the loss of the affinity matrix in the passing process. Furthermore, an alternating minimization optimization method is designed in this paper to solve the optimization problem. Experimental results on seven underlying datasets demonstrate the efficiency and superiority of the proposed method.