Multiclass Feature Selection Via Kernel Parameter Optimization

This paper considers feature selection in a multiclass classification scenario where the goal is to determine a subset of available features which is most discriminative and informative for all the classes simultaneously. Based on the data distributions of classes in the feature space, this paper first presents a model selection criterion named multiclass kernel polarization (MKP) to evaluate the goodness of a kernel in multiclass classification scenario, and then optimizes the scale factors assigned to each feature in a kernel by maximizing this criterion to identify the more relevant features. The proposed method is demonstrated with two UCI machine learning benchmark examples.