A robust ensemble feature selector based on rank aggregation for developing new VO2max prediction models using support vector machines

Abstract

This paper proposes a new ensemble feature selector, called the majority voting feature selector (MVFS),for developing new maximal oxygen uptake (VO2max) prediction models using a support vector machine (SVM). Theapproach is based on rank aggregation, which meaningfully utilizes the correlation among the relevance ranks of predictorvariables given by three state-of-the-art feature selectors: Relief-F, minimum redundancy maximum relevance (mRMR),and maximum likelihood feature selection (MLFS). By applying the SVM combined with MVFS on a self-created datasetcontaining maximal and submaximal exercise data from 185 college students, several new hybrid VO2max predictionmodels have been created. To compare the performance of the proposed ensemble approach on prediction of VO2max,SVM-based models with individual combinations of Relief-F, mRMR, and MLFS as well as with other alternativeensemble feature selectors from the literature have also been developed. The results reveal that MVFS outperformsother individual and ensemble feature selectors and yields up to 8.76% increment and 11.15% decrement rates in multiplecorrelation coefficients (Rs) and root mean square errors (RMSEs), respectively. Furthermore, in addition to reconfirmingthe relevance of sex, age, and maximal heart rate in predicting VO2max, which were previously reported in the literature,it is revealed that submaximal heart rates and exercise times at 1.5-mile distance are two further discriminative predictorsof VO2max. The results have also been compared to those obtained by a general regression neural network and singledecision tree combined with MVFS, and it is shown that the SVM exhibits much better performance than other methodsfor prediction of VO2max.