A Theory of Actionable Data Mining with Application to Semiconductor Manufacturing Control

Accurate and timely prediction of a manufacturing process yield and flow times is often desired as a means of reducing the overall production costs. To this end, we develop in this paper a new decisiontheoretic classification framework and apply it to a real-world semiconductor wafer manufacturing line, which suffers from constant variations in the characteristics of the chip manufacturing process. The decision-theoretic framework is based on a model for evaluating classifiers in terms of their value in decision-making. Recognizing that in many practical applications the values of the class probabilities as well as payoffs are neither static nor known exactly, we present a precise condition under which one classifier ‘dominates’ another classifier (i.e., achieves higher payoff), regardless of payoff or class distribution information. Building on the decision-theoretic model, we propose two robust ensemble classification methods that construct composite classifiers which are at least as good as any of the existing component classifiers for all possible payoff functions and class distributions. We show how these two robust ensemble classifiers are put into practice by developing decision rules for effectively monitoring and controlling the real-world semiconductor wafer fabrication line under study.