An empirical analysis of the benefit of decision tree size biases as a function of concept distribution
The results reported here empirically show the benefit of decision tree size biases as a function of concept distribution. First, it is shown how concept distribution complexity (the number of internal nodes in the smallest decision tree consistent with the example space) affects the benefit of minimum size and maximum size decision tree biases. Second, a policy is described that defines what a learner should do given knowledge of the complexity of the distribution of concepts. Third, explanations for why the distribution of concepts seen in practice is amenable to the minimum size decision tree bias are given and evaluated empirically.