Bradley Novic, PhD, Principal Consultant, PhaseTwo Analytics
Mining manufacturing data can be a perilous endeavor from data assembly to analysis to interpretation and implementation of results. What sets data mining in manufacturing apart from data mining in, say, marketing or finance, is that prediction alone is not good enough in manufacturing. In manufacturing the end game is improved control, and that requires getting to root cause. Decision trees are a wonderful tool for discovery, but left to their own devices can select variables that, although predictive, make no sense to a process engineer from a control standpoint. This requires a more careful approach to tree building that is afforded by the stepwise approach provided in the JMP Partitioning platform and unavailable in most other decision tree software programs. This paper will provide examples of data mining successes using JMP in manufacturing and issues in mining manufacturing data. It will also demonstrate the integration of expert knowledge into the JMP partitioning process, to deal with the issue of correlation in predictor variables and provide rules that permit validation and deployment of results.