Feature Evalution with Measures of Probabilistic Dependence

In this paper, measures of probabilistic dependence are derived from distance measures and are applied to feature evaluation in pattern recognition. The main properties of the measures are derived and are discussed in their application to feature-class dependency. Relations between the measures and error probability are derived. Experiments using feature subsets extracted from Munson's hand-printed data are performed to compare the feature-evaluating capabilities of the measures both relative to each other and relative to error probability.