Using the ADTree for feature reduction through knowledge discovery

There is a chicken-and-egg problem in classification whereby a good classifier is required to test the efficacy of the features, yet a good feature set is required to generate a good classifier. When the salient features are unknown, an extremely large set of features is used to train the classifier in hopes of obtaining accurate classification results. This research proposes the use of a special class of decision tree called the alternating decision tree or ADTree to answer two questions in knowledge discovery in order to effectively select a salient feature set: When using a particular feature extraction algorithm, which of the features is able to distinguish between the different classes? And how do they work?