Measuring Features Strength in Probabilistic Classification

Probabilistic classifiers output a probability of an input being a member of each of the possible classes, given some of its feature values, selecting most probable class as predicted class. We introduce and compare different measures of the feature strength in probabilistic confidence-weigthed classification models. For that, we follow two approaches: one based on conditional probability tables of the classification variable with respect to each feature, using different statistical distances and a correction parameter, and the second one based on accuracy in predicting classification from evidences on each isolated feature. On a case study, we compute these feature strength measures and rank features attending to them, comparing results.