A Framework For Ensemble Learning Based Heterogeneous Extreme Learning Machines

Neural networks(NNs)are one of machine learning techniques.NN suffers from time-consuming learning,local minimal and human interfere.However,recently Extreme learning machine(ELM)has gained popularity for solving classification problems.ELM is a single-hidden layer feed-forward network(SLFN)extension.It performs well in managing some problems due to its fast learning speed and performance generalization.Moreover,ensemble learning offers an inexpensive alternative due to its performance optimization.Ensemble learning is a machine learning process to get better prediction performance by strategically combining the predictions from multiple learning algo-rithms.Ensembles are known to reduce the risk of selecting the wrong model by ag-gregating all candidate models.Ensembles are known to be more accurate than single models.Accuracy has been identified as an important factor in explaining the success of ensembles.Several techniques have been proposed to improve ensemble accuracy.But till now no perfect one is proposed.The focus of this research is on how to create accurate ensembles based ELM in the context of classification to deal with supervised data,noisy data,imbalanced data,and semi-supervised data.To deal with the mentioned above issues we propose a framework for heterogeneous ELM ensemble.To tackle noisy data we use the correntropy to achieve insensitive performance to outliers,while implementing negative correlation learning(NCL)to en-hance diversity among the ensemble.The proposed heterogeneous ensemble of ELMs(HE2LM)for classification has different ELM algorithms,including regularized ELM(RELM),kernel ELM(KELM),and L2-norm-optimized ELM(ELML2).To deal with imbalanced data we propose new diverse AdaBoost ensemble based ELM(AELME)for binary and multiclass data classification.To deal with semi-supervised issue we introduce a multi-kernel semi-supervised ELM(MKSSELM)algorithm.It is more flex-ible to deal with discrete data from various sources.It matches diverse information from disparate sources and it shows distinction among the data.Instead of using one kernel,we optimize both ELM structural parameters and kernel combination weights.The optimization process accomplished by commanding an L1-norm as a regulation term.Meanwhile,we use a non-negative constraint on the kernel combination weights.We apply the ensemble algorithm to large datasets.Moreover,we apply it to three case studies.Namely,functional magnetic resonance imaging,speaker recognition,and covertype data.