Ron Kohavi and Daniel Sommerfield
Business users and analysts commonly use spreadsheets and 2D plots to analyze and understand their data. On-line Analytical Processing (OLAP) provides these users with added flexibility in pivoting data around different attributes and drilling up and down the multi-dimensional cube of aggregations. Machine learning researchers, however, have concentrated on hypothesis spaces that are foreign to most users: hyperplanes (Perceptrons), neural networks, Bayesian networks, decision trees, nearest neighbors, etc. In this paper we advocate the use of decision table classifiers that are easy for line-of-business users to understand. We describe several variants of algorithms for learning decision tables, compare their performance, and describe a visualization mechanism that we have implemented in MineSet. The performance of decision tables is comparable to other known algorithms, such as C4.5/C5.0, yet the resulting classifiers use fewer attributes and are more comprehensible.