Bias and stability of single variable classifiers for feature ranking and selection
Development of a feature ranking method based upon the discriminative power of features and unbiased towards classifiers is of interest. We have studied a consensus feature ranking method, based on multiple classifiers, and have shown its superiority to well known statistical ranking methods. In a target environment such as a medical dataset, missing values and an unbalanced distribution of data must be taken into consideration in the ranking and evaluation phases in order to legitimately apply a feature ranking method. In a comparison study, a Performance Index (PI) is proposed that takes into account both the number of features and the number of samples involved in the classification.