Heuristics for efficient classification
Kathryn L. Fraughnaugh, Holly Zullo, Louis Anthony Cox, Joel L. Ryan · 2002
The objective of this project was to develop, implement and test heuristics to find an inspection strategy for classification. Given a set of attribute values for deciding class membership, prior statistical information about the relative frequencies of attribute values, and costs of inspection of attribute values, what is an optimal sequential inspection strategy for determining the class of some object? This paper introduces a simple dynamic rule for classification that is easily represented. Simple variation of components of the rule lead to a wide search of the set of all decision trees. This is borne out by results in which the outcome of a random search of all decision trees is compared with that of a random search of decision trees that can be represented by our rule. The construction of our rule is flexible, and could easily be varied to encompass important components of classification problems that differ from ours. The results show that the tabu searches are very effective, delivering a near optimal strategy that a classifier can use repeatedly with near minimum expected long run inspection costs.>