Robust self-tuning multi-view clustering

Previous methods of multi-view clustering focused on the improvement of clustering effectiveness by detecting common information of all views and individual information for every view, but they ignore the following issues, i.e., the initialization sensitivity, the cluster number determination, and the influence of outliers. However, either single-view clustering or multi-view clustering often suffers from above issues. In this paper, we propose a robust self-tuning multi-view clustering to introduce a sum-of-norm loss function to explore the issue of initialization sensitivity, design a sum-of-norm regularization to automatically determine the cluster number, and employ robust statistics techniques to reduce influence of outliers. Furthermore, we propose an effective alternating optimization method to solve the resulting objective function and then theoretically prove its convergence. Experimental results on both synthetic and real data sets demonstrated that our proposed multi-view clustering method outperformed the state-of-the-art clustering methods, in terms of four clustering evaluation metrics.