Lower-bound Estimate for Cost-sensitive Decision Trees

While an extensive body of literature investigates problems of decision trees growing, just a few study lower-bound estimates for the expected classification cost of decision trees, especially for varying costs of tests. In this paper the new lower-bound estimate is proposed. Computation of the estimate is reduced to solving a series of set-covering problems. Computational complexity and other properties of the lower-bound estimate are investigated. The top-down algorithm of tree construction based on the proposed estimate is tested against several popular greedy cost-sensitive heuristics on a range of standard data sets from UCI Machine Learning Repository.