Automatically Balancing Accuracy and Comprehensibility
Rikard König · 2005
Onespecific problem, whenperforming predictive modeling, isthetradeoff between accuracy andcomprehensibility. Whencomprehensible models are required thisnormally rulesouthigh-accuracy techniques like neural networks andcommittee machines. Therefore, anautomated choice ofastandard technique, knowntogenerally produce sufficiently accurate and comprehensible models, would beofgreat value. Inthis paperitisargued that this requirement ismetbyan ensemble ofclassifers, followed byrule extraction. The proposed technique isdemonstrated, using anensemble ofcommonclassifiers andourrule extraction algorithm G-REX, on17publicly available datasets. Theresults presented demonstrate thatthesuggested technique performs verywell. Morespecifically, theensemble clearly outperforms theindividual classifiers regarding accuracy, while theextracted models haveaccuracy similar totheindividual classifiers. Theextracted models are,however, significantly more compactthan corresponding models created directly fromthedata set using thestandard tool CART;thus providing higher comprehensibility.