Generating Accurate Rule Sets Without Global Optimization

The two dominant schemes for rule learning C and RIPPER both operate in two stages First they induce an initial rule set and then they re ne it using a rather com plex optimization stage that discards C or adjusts RIPPER individual rules to make them work better together In con trast this paper shows how good rule sets can be learned one rule at a time with out any need for global optimization We present an algorithm for inferring rules by repeatedly generating partial decision trees thus combining the two major paradigms for rule generation creating rules from de cision trees and the separate and conquer rule learning technique The algorithm is straightforward and elegant despite this ex periments on standard datasets show that it produces rule sets that are as accurate as and of similar size to those generated by C and more accurate than RIPPER s More over it operates e ciently and because it avoids postprocessing does not su er the ex tremely slow performance on pathological ex ample sets for which the C method has been criticized