Research On Data Processing And Sample Classification Methods Based On Extreme Learning Machin

With the promotion and popularization of Machine Learning,Extreme Learning Machine(ELM)has been widely concerned and developed.Recently,experts and scholars have proposed many methods to improve the robustness of ELM.At the same time,with the rapid development of computer hardware and the increasingly mature Internet technology,the idea of ELM integration has also begun to enter the field of vision of researchers.How to improve the stability of ELM and how to further improve the classification performance of ELM while maintaining the robustness of ELM have become two important directions of ELM research.This thesis studies ELM in the above two directions,and uses the improved model to complete specific classification tasks.Compared with some excellent ELM methods at the present stage,the improved model proposed in this thesis has more ideal classification performance.The specific research contents are as follows:(1)A kernel risk-sensitive loss hyper-graph regularization ELM model based on sparrow search algorithm is proposed.In the traditional ELM method,the input weights and deviations between the input layer and the hidden layer are generated randomly in the training process,which hinders the further improvement of the ELM classification performance.To solve this problem,this thesis introduced sparrow search algorithm to optimize the input weight and deviation of ELM,so as to improve the classification performance of ELM model.In addition,the adaptive function of sparrow search algorithm is redesigned in this thesis.Finally,the corresponding experiments are carried out,and the experimental results show that the new model has better classification performance.(2)An integrated ELM model based on q-Gaussian kernel risk-sensitive loss and voting mechanism is proposed.The new model is inspired by the idea of parameterization of function family and integration.The parameterization of function family enables the original data to obtain a better feature space representation and enhances the generalization ability of the model.In addition,the traditional voting mechanism is introduced for integration.Compared with a single ELM classifier,the integrated model has better classification stability.The experimental results show that the new model can improve the classification performance and has certain competitiveness in the classification stability of the model.(3)An integrated ELM model with variable sample weights is proposed.In most traditional ELM,due to its default that all samples are of the same importance,ELM lacks sensitivity to a particular category.Therefore,this thesis designed a quantization method of sample weight,and proposed a new integrated ELM model on this basis.The new model improved the data processing process and used divide-and-conquer strategy in the classification prediction process.In this thesis,the model is applied to two learning scenarios of supervised learning and semisupervised learning.The experimental results show that compared with other excellent ELM methods,the classification prediction ability of the new model is more prominent.In a word,this thesis proposed three ELM models from the perspectives of internal structure,stability and integration of ELM,and designed corresponding experiments to compare with existing excellent ELM methods.The results show that the model proposed in this thesis has certain competitiveness in classification performance or classification stability,and can complete specific classification prediction tasks.