Search

Feature selection or elimination methods have been widely used to mine high-dimensional data from many fields such as biomedicine. However, classical feature selection methods usually focus on class-balanced and numerical data, and the classification performance is poor when processing class-imbalanced or mixed-type data. Moreover, existing studies hardly consider the co-occurrence of both class-imbalance and mixture of numerical and nominal features. In this paper, we propose an adaptive loss backward feature elimination method for class-imbalanced and mixed-type data (FEIM). In FEIM, a technique to convert and decompose nominal features is employed to process mixed-type data. Meanwhile, an adaptive loss function is proposed to make the classifier pay more attention to the minority class. In experiments, the effectiveness and stability of FEIM are synthetically evaluated against seven medical diagnostic datasets in terms of the measurements of F measurement and precision, and the results of feature selection are analyzed in detail. The experimental results show that, compared with three common methods (Recursive Feature Elimination SVM (RFE-SVM), Relief, and Holdout Algorithm for Backward Feature Elimination (HO-BFE)), FEIM can achieve a higher classification accuracy and is more stable when the parameters are changed. On the other hand, the feature selection results of FEIM are more meaningful by processing class-imbalanced and mixed-type data. Therefore, FEIM is a competitive and effective method to process feature elimination for class-imbalanced and mixed-type biomedical data for in medical diagnosis. The supporting information is available in the electronic supplementary material. • A mixed-type feature selection algorithms called FEIM is proposed for imbalanced data with both numerical and nominal features. • An adaptive loss function is designed for class-imbalance learning. • FEIM is more stable when the model parameters are changed compared with several classical methods. [ABSTRACT FROM AUTHOR]

This paper is the generalization of weight-fused elastic net (Fu and Xu, 2012), which performs group variable selection by combining weight-fused LASSO(wfLasso) and elastic net (Zou and Hastie, 2005) penalties. In this study, the elastic net penalty is replaced by adaptive elastic net penalty (AdaEnet) (Zou and Zhang, 2009), and a new group variable selection algorithm with oracle property (Fan and Li, 2001; Zou, 2006) is obtained. [ABSTRACT FROM PUBLISHER]

In this paper a novel wavelength region selection algorithm, called elastic net grouping variable selection combined with partial least squares regression (EN-PLSR), is proposed for multi-component spectral data analysis. The EN-PLSR algorithm can automatically select successive strongly correlated prediction variable groups related to the response variable using two steps. First, a portion of the correlated predictors are selected and divided into subgroups by means of the grouping effect of elastic net estimation. Then, a recursive leave-one-group-out strategy is employed to further shrink the variable groups in terms of the root mean square error of cross-validation (RMSECV) criterion. The performance of the algorithm with real near-infrared (NIR) spectroscopic data sets shows that the EN-PLSR algorithm is competitive with full-spectrum PLS and moving window partial least squares (MWPLS) regression methods and it is suitable for use with strongly correlated spectroscopic data. [ABSTRACT FROM AUTHOR]