Building decision trees with constraints

Classification is an important problem in data mining. Given a database of records, each with a class label, a classifier generates a concise and meaningful description for each class that can be used to classify subsequent records. A number of popular classifiers construct decision trees to generate class models. Frequently, however, the constructed trees are complex with hundreds of nodes and thus difficult to comprehend, a fact that calls into question an often-cited benefit that decision trees are easy to interpret. In this paper, we address the problem of constructing "simple" decision trees with few nodes that are easy for humans to interpret. By permitting users to specify constraints on tree size or accuracy, and then building the "best" tree that satisfies the constraints, we ensure that the final tree is both easy to understand and has good accuracy. We develop novel branch-and-bound algorithms for pushing the constraints into the building phase of classifiers, and pruning early tree nodes that cannot possibly satisfy the constraints. Our experimental results with real-life and synthetic data sets demonstrate that significant performance speedups and reductions in the number of nodes expanded can be achieved as a result of incorporating knowledge of the constraints into the building step as opposed to applying the constraints after the entire tree is built.