Abstract: Traditional feature selection algorithm used a single scalar metric such that it might become difficult to achieve a trade-off between generalization performance and dimension reduction at the same time. A new feature selection algorithm called LS-SVM-FSC was proposed to circumvent this shortcoming. The kernel-based least squares support vector machines was used to train a set of binary classifiers on each single feature and a kind of new fuzzy membership function was used to obtain fuzzy membership value of each pattern belonging to its class. Based on a new fuzzy supplementary criterion, the features with minimal redundancy and maximal relevance was selected. Experiments indicated that the proposed algorithm had high classification accuracy and strong dimension reduction capability on nine datasets. In particular, it still kept fast learning speed for high-dimensional datasets, in contrast to other ten feature selection methods and seven degree determination methods.

Key words: feature selection, fuzzy supplementary criterion, least squares support vector machines, classification, fuzzy membership degree function