Representation Design and Brute-force Induction in a Boeing Manufacturing Domain

We applied inductive classi cation techniques to data collected in a Boeing plant with the goal of uncovering possible aws in the manufacturing process. This application led us to explore two aspects of classical decision-tree induction: Preand Post-processing: much of our e ort was focused on the pre-processing of raw data to make it suitable for induction and the post-processing of learned rules to make them useful to factory personnel. Brute-force Induction: in contrast with standard methods, which perform a greedy search of the space of decision trees, we formulated an algorithm that conducts an exhaustive, depth-bounded search for accurate predictive rules. We demonstrate the e cacy of our approach with speci c examples of learned rules and by quantitative comparisons with decision-tree algorithms (C4 and CART). This research was funded in part by a Boeing Computer Services contract with the University of Washington, by O ce of Naval Research Grant 92-J-1946, and by National Science Foundation Grants IRI9211045 and IRI-9357772 (NYI Award to Etzioni). Richard Segal is supported, in part, by a GTE fellowship. The C routines, implementing the Brute and Gold-digger algorithms, are available by sending mail to [email protected]. We are grateful to Wray Buntine for distributing his IND package, which contains re-implementations of C4 and CART. IND greatly facilitated our research. Thanks are due to Ruth Etzioni for her expert advice on statistical testing, to Usama Fayyad and Je Schlimmer for helpful suggestions, and to Mike Barley for many fruitful discussions and reviews of drafts of this paper. Finally, we acknowledge the other members of the Boeing project: Mike Healy, Dave Newman, and Carl Pearson.