Pinball loss based extreme learning machines
Kuaini Wang, Xiaoshuai Ding · IOP Conference Series Materials Science and Engineering · 2019
Abstract Extreme Learning Machine (ELM) is a novel machine learning method by training single hidden layer feedforward neural network. It employs squared loss function to minimize the mean squares error, which is sensitive to noises and outliers. In this paper, pinball loss function with quantile error is introduced into ELM in order to improve the robustness of ELM. An ELM model based on squared pinball loss function (SPELM) and an ELM model based on pinball loss function (PELM) are proposed. The corresponding optimization problem are solved by iterative reweighted algorithm. Three simulated datasets and nine Benchmark datasets are used to verify the validity of the proposed models. It is concluded that the proposed SPELM and PELM are superior to other comparisons, especially for datasets containing larger proportion of outliers.