Weight Learning In Weighted ELM Classification Model Based On Genetic Algorithms

In the field of machine learning and data mining,introducing the concept of class error cost into the design of classifier calls cost sensitive learning.In the problem of cost sensitive classification,we often suppose to have a known cost matrix in which each element represents the cost of mistakenly classifying an object from one class into another.Weighted least squares is a typical method for dealing with cost sensitive classification problems.Allocating reasonable weights to different categories will greatly improve the classification ability of classification models.However,in practical problems,we only know that these weights depend on the cost matrix,but rarely see how to determine these weights according to the cost matrix.This paper is to study a weighted least square model of ELM,which determines the weight of training samples from the cost matrix based on genetic algorithm.The paper mainly completed the following work:(1)Weighted least squares model of ELM is studied,which is based on genetic algorithm to determine the weight of training samples from the cost matrix.Cost sensitive classifier can deal with problem of unbalanced classification and can greatly reduce cost of average error classification.Different weights have different influence on the classifier.Allocating reasonable weights to different categories will greatly improve the classification ability of classifier.It is found from the experiment that the genetic algorithm can solve the problem in this paper.For different data sets,total cost of weighted least squares model is approximately minimum and we get the weights of training samples corresponding to total cost of weighted least squares model.(2)The relationship between cost matrix and weight in weighted least squares model is studied.The experimental results show that,for two and three categories,the higher the value of sum up certain column in the cost matrix is the greater weight of a certain class.The higher the value of sum up certain column in the cost matrix,the greater cost of other classes being classified into certain class.The greater weight of a certain class,that means it let the classification boundary be far away from certain class,and increases the survival space of certain class.It also increases the probability that allocated samples fall to certain class.At the same time,randomness of genetic algorithm has little effect on the conclusion.(3)Average total cost of weighted least squares model is compared with total cost of cost sensitive ELM.The experiment shows size of total cost as an indicator,average total cost of weighted least squares model is smaller than total cost of cost sensitive ELM.(4)Average total cost of weighted least squares model is compared with total cost of cost sensitive naive Bayes model.Experimental results show that size of total cost as an indicator,average total cost of weighted least squares model is better than total cost of cost sensitive naive Bayes model.Model analysis and experimental results show effectiveness of method.Further comparison shows that size of total cost as an indicator,weighted least squares model is far superior to existing cost sensitive ELM and cost sensitive naive Bayes model.At the same time,the higher cost of adding a digit to certain column in the cost matrix,the greater weight of certain class.