Spectra data classification with kernel extreme learning machine

Abstract

One of the most important issues in chemometrics is how to design a robust classifier for spectra data with high classification accuracy. Extreme learning machine (ELM) has proved to be an effective method for numerous pattern recognition problems. However, the randomness of ELM’s hidden layer nodes setting would give rise to the oscillation of the hidden layer output matrix, which may bring down the stability of this model. This paper proposes a framework for spectra data classification using the kernel extreme learning machine (KELM) that can circumvent the calculation of hidden layer outputs and encode it in a kernel matrix inherently. Moreover, KELM doesn’t need to specify the hidden layer feature mapping function, and the number of hidden nodes doesn’t need to be given, which may avoid an extensive mapping calculation. The proposed framework contains two stages: spectral preprocessing and dimensionality reduction, as well as classification implementation with KELM. The experimental results on four datasets reveal that the radial basis function (RBF) should be a priority candidate kernel function for this kind of applications, and the suggestion of parameter selection is also given experimentally. Furthermore, the classification accuracy of KELM is superior to ELM in most cases, and KELM is significantly faster than ELM when the number of the training samples is relatively small, which is a common case for spectra data in chemometrics. Therefore, KELM may be a promising method for rapid spectra data classification.