Mixed structure low-rank representation for multi-view subspace clustering

Multi-view clustering method utilizes the diversity of multi-view information to access better clustering results than a single view. Most existing multi-view clustering methods do not take full advantage of the diversity of information views, which makes the affinity matrix insufficiently clear and accurate to precisely describe the potential structure of multi-view data, resulting in poor clustering results. To solve the above problems, mixed structure low-rank representation (MSLRR) for multi-view subspace clustering and its kernel version (ker-MSLRR) are proposed in this paper. The mixed low-rank structure algorithm takes the multi-view data after the feature concatenation as input and then uses the nested mixed structure of least squares regression (LSR) and low-rank representation (LRR) as the unified model to effectively reduce the noise of the affinity matrix. In addition, to effectively deal with nonlinear data, the kernel method ker-MSLRR based on MSLRR is proposed, which improves the processing ability of processing nonlinear data. The experimental results of five real datasets demonstrate that the proposed methods have better clustering performance than other existing methods.