A New Fuzzy Extreme Learning Machine for Regression Problems with Outliers or Noises

Abstract

Extreme Learning Machine (ELM), recently proposed by Huang et al., has attracted much attention from more and more researchers in the machine learning and data mining community, and has shown similar or better generalization performance with dramatically reduced training time than Support Vector Machines (SVM). In ELM, it is implicitly assumed that all samples in training datasets share the same importance. Therefore, when it comes to datasets with outliers or noises, like SVM, ELM may produce suboptimal regression models due to overfitting. In this paper, by equipping ELM with the fuzzy concept, we propose a novelty approach called New Fuzzy ELM (NF-ELM) to deal with the above problem. In NF-ELM, firstly, different training samples are assigned with different fuzzy-membership values based on their degree of being outliers or noises. Secondly, these membership values are incorporated into the ELM algorithm to make it less sensitive to outliers or noises. The performance of the proposed NF-ELM algorithm is evaluated on three artificial datasets and thirteen real-world benchmark function approximation problems. The results indicate that the proposed NF-ELM algorithm achieves better predictive accuracy in most cases than ELM and SVM does.