Abstract: A computer implemented system, two-step method and computer program product which improves the operations of multi-feature extraction and which efficiently develops classification rules from a large training database. Specifically, given a large training set of data tuples, the first phase, called the feature identification phase, identifies features, which have good power in separating data tuples, based on a subset of the training set. A feature that has a good power in correlating data tuples into groups is said to have a good discriminating power. Since the feature identification phase is performed on a subset of the training set, processing costs are minimized. Limiting this phase to the identification of features having good discriminating power ensures that the use of a subset of the training set does not adversely affect the validity of the conclusions drawn therefrom.