LAUSR

Abstract Though many successful methods have been proposed for supervised learning tasks, such as support vector machines and extreme learning machines (ELM), it is still an open problem to extend the successful supervised learning methods to unsupervised learning tasks and obtain better results. In this paper, we propose to extend the ELM to an unsupervised learning version and propose a clustering method based on ELM (CM-ELM) for both binary class and multiple class problems, which aims to find a labeling that would yield an optimal ELM classifier. In the ELM feature space, we propose to combine the Gaussian hidden nodes and sigmoid hidden nodes in the hidden layer to combine their advantages. Then we propose to adopt the alternative direction method to solve the non-convex problems in CM-ELM simply. Furthermore, in order to make the results of the non-convex problems robust and satisfactory, we propose to initial the labels with cluster ensemble methods. Experiments on the artificial and benchmark data sets show that the CM-ELM is competitive to the state-of-the-art clustering methods.