Least Squares Twin Extreme Learning Machine for Pattern Classification

In this paper, we have proposed Least Squares Twin Extreme Learning Machine (LSTELM) for pattern classification. The proposed LSTELM formulation solves Extreme Learning Machine (ELM) problem in twin framework. It owns better generalization ability than standard ELM (Huang et al. in IEEE Trans Syst Man Cybern: Part B (Cybern) 42(2):513–529 (2012) [1]) by finding two nonparallel hyperplanes in ELM random feature space which passes through the origin. Unlike Twin Extreme Learning Machine (TELM) (Wan et al. in Neurocomputing 260:235–244 (2017) [2]), it requires less computation time for training the model as proposed LSTELM can be efficiently solved by solving only two systems of linear equations. The proposed LSTELM formulation combines the benefits of the TWSVM (Jayadeva et al. in IEEE Trans Pattern Anal Mach Intell 29(5):905–910 (2007) [3]) and standard ELM (Huang et al. in IEEE Trans Syst Man Cybern: Part B (Cybern) 42(2):513–529 (2012) [1]) in true sense. Experimental results on several UCI benchmark datasets (Bache et al. in UCI Machine Learning Repository (2013) [4]) show that the proposed LSTELM is well generalizable than standard ELM and take less training time compared to TELM model (Wan et al. in Neurocomputing 260:235–244 (2017) [2]).