An adaptive feature selection method for multi-class classification

A vast variety of feature selection methods have been proposed according to different metrics, such as information gain, entropy, chi-square test, t-test. Yet when applied to multi-class classification task, these methods generally suffer the “siren pitfall” of a surplus of predictive features for some classes while lack of predictive features for the remaining classes. A number of solutions to the “siren pitfall” have been proposed, yet there are still problems with these methods. For example, the selected features by “randomized feature set” method are not re-configurable; the “rand-robin” method and the “round-robin” method may miss some important features whenever the one v.s. others binary partition does not work; the computation cost of the wrapper's method is rather high. In this paper, we propose an adaptive feature selection method for multi-class classification task. With our method, the “siren pitfall” could be avoided, the selected features could be reproduced, the feature selection scheme does not rely on any prior knowledge, and the corresponding computation cost is low. Experimental results indicate the effectiveness of our adaptive feature selection method.