| Early diagnosis of cancer on the genetic level can effectively improve the cure rate of patients.However,cancer gene expression data are confronted with the characteristics of high dimensionality,small sample size,high noise level,and the class imbalance,which bring up great challenge to classification.At present,using ensemble algorithm to analyze gene expression data is an important research topic in the field of bioinformatics.However,as a result of the characteristics of gene expression data,as well as the ensemble learning algorithm itself,we tend to have difficulty in analysing gene expression data by ensemble learning.For example,it is difficult to balance the diversity and accuracy of base learners in selective ensemble algorithm and hard to optimize the parameters of each base learner in the ensemble algorithm.Aiming at solving the above problems,this paper has the following contributions to the literature:(1)The diversity between the base classifiers and the accuracy of each single base classifier were two important factors that determine the generalization performance of ensemble system.In view of the difficulty in balancing the diversity and accuracy,a selective ensemble algorithm for gene expression data based on the diversity and accuracy of the weight harmonic average(D-A-WHA)was proposed.The kernel extreme learning machine was used as the base classifiers,and the diversity and accuracy of the base classifiers was adjusted by the D-A-WHA measure.Finally,a set of classifiers with high diversity and accuracy was selected as the base classifiers of selective ensemble.Compared with traditional Bagging,Adaboost and other ensemble algorithms,the classification accuracy of D-A-WHA selective ensemble algorithm was improved by 1%-3% on multiple gene expression data.(2)A differential evolution based on cost-sensitive Stacking ensemble(DE-CStacking)for cancer gene expression data classification was proposed.Random Forest,K-Nearest Neighbor,and Naive Bayes were used as lower-level learners(level-0)of stacking ensemble,and the cost-sensitive Support Vector Machine was used as the high-level learner(level-1).The original feature sets and the output class probabilities of the lower-level learners are all used as the input of the high-level learner.The parameters of these learners are optimized by differential evolution.Compared to other traditional Stacking ensemble algorithm,the AUC value of DE-CStacking algorithm was improved by 2%-7% on multiple gene expression data. |
