Multi-view Clustering Based on Low-rank Representation and Adaptive Graph Learning

The multi-view algorithm based on graph learning pays attention to the manifold structure of data and shows good performance in clustering task. However, multi-view data usually contains noise, which reduces the robustness of the multi-view clustering algorithm. Besides, any single local information cannot adequately express the whole frame perfectly. Graph learning method often ignores the global structure of data, resulting in suboptimal clustering effect. In order to address the above problems, we propose a novel multi-view clustering model, namely multi-view clustering based on low-rank representation and adaptive graph learning (LRAGL). The noise and outliers in the original data are considered when constructing the graph and the adaptive learning graphs are employed to describe the relationship between samples. Specifically, LRAGL enjoys the following advantages: (1) The graph constructed on the low-rank representation coefficients after filtering out the noise can more accurately reveal the relationship between the samples; (2) Both the global structure (low-rank constraints) and the local structure (adaptive neighbors learning) in the multi-view data are captured; (3) The filtering of noise and the construction of the similarity graph of each view data are integrated into a framework to obtain the overall optimal solution; LRAGL model can be optimized efficiently by utilizing the augmented Lagrangian multiplier with Alternating Direction Minimization Method (ADMM). Extensive experimental results on six benchmark datasets verify the superiority of the proposed method in clustering.