Extreme learning machine for regression based on condition number and variance decomposition ratio

Abstract

The extreme learning machine (ELM) is a novel single hidden layer feedforward neural network. Compared with traditional neural network algorithm, ELM has the advantages of fast learning speed and good generalization performance. However, there are still some shortages that restrict the further development of ELM, such as the perturbation and multicollinearity in the linear model. To the adverse effects caused by the perturbation and the multicollinearity, this paper proposes ELM based on condition number and variance decomposition ratio (CVELM) for regression, which separates the interference terms in the model by condition number and variance decomposition ratio, and then manipulate the interference items with weighted. Finally, the output layer weight is calculated by the least square method. The proposed algorithm can not only get good stability of the algorithm, but also reduce the impact on the non-interference items when dealing with the interference terms. The regression experiments on several datasets show that the proposed method owns a good generalization performance and stability.