Fast Effective Rule Induction-英文文献.pdf

From Machine Learning Pro ceedings of the Twelfth International Conference ML Fast Eective Rule Induction William W Cohen ATT Bell Laboratories Mountain Avenue Murray Hill NJ wcohenresearchattcom Abstract this problem is of critical importance The goal of this paper is to develop propositional rule learning algo Many existing rule learning systems are rithmsthat perform eciently on large noisy datasets computationally expensive on large noisy that extend naturally to rstorder representations datasets In this paper we evaluate the andthatarecompetitiveingeneralizationperformance recentlyproposed rule learning algorithm with more mature symbolic learning methods such IREP on a large and diverse collection of as decision trees The end product of this eort is benchmark problems We show that while the algorithm RIPPERk which is competitive with IREP is extremely ecient it frequently Crules with respect to error rates scales nearly lin gives error rates higher than those of C early with the number of training examples and can and Crules We then propose a num eciently process noisy datasets containing hundreds ber of modications resulting in an algo of thousands of examples rithm RIPPERk that is very competitive with Crules with respect to error rates PREVIOUS WORK but much more ecient on large samples RIPPERk obtains error rates lower than or equivalent to Crules on of bench COMPLEXITY OF RULE P