Optimization extreme learning machine with ν regularization

The problem of choosing error penalty parameter C for optimization extreme learning machine (OELM) is that it can take any positive value for different applications and it is therefore hard to choose correctly. In this paper, we reformulated OELM to take a new regularization parameter nu (nu-OELM) which is inspired by Scholkopf et al. The regularization in terms of nu is bounded between 0 and 1, and is easier to interpret as compared to C. This paper shows that: (1) nu-OELM and nu-SVM have similar dual optimization formulation, but nu-OELM has less optimization constraints due to its special capability of class separation and (2) experiment results on both artificial and real binary classification problems show that nu-OELM tends to achieve better generalization performance than nu-SVM, OELM and other popular machine learning approaches, and it is computationally efficient on high dimension data sets. Additionally, the optimal parameter nu in nu-OELM can be easily selected from few candidates. (C) 2017 Elsevier B.V. All rights reserved.