High-Performance Commercial Data Mining: A Multistrategy Machine Learning Application

摘要

We present an application of inductive concept learning and interactive visualization techniques to a large-scale commercial data mining project. This paper focuses on design and configuration of high-level optimization systems (wrappers) for relevance determination and constructive induction, and on integrating these wrappers with elicited knowledge on attribute relevance and synthesis. In particular, we discuss decision support issues for the application (cost prediction for automobile insurance markets in several states) and report experiments using D2K, a Java-based visual programing system for data mining and information visualization, and several commercial and research tools. We describe exploratory clustering, descriptive statistics, and supervised decision tree learning in this application, focusing on a parallel genetic algorithm (GA) system, Jenesis, which is used to implement relevance determination (attribute subset selection). Deployed on several high-performance network-of-workstation systems (Beowulf clusters), Jenesis achieves a linear speedup, due to a high degree of task parallelism. Its test set accuracy is significantly higher than that of decision tree inducers alone and is comparable to that of the best extant search-space based wrappers.