Resource constraints are commonly found in classification tasks. For example, there could be a budget limit on implementation and a deadline for finishing the classification task. Applying the top-down approach for tree induction in this situation may have significant drawbacks. In particular, it is difficult, especially in an early stage of tree induction, to assess an attributes contribution to improving the total implementation cost and its impact on attribute selection in later stages because of the deadline constraint. To address this problem, we propose an innovative algorithm, namely, the Cost-Sensitive Associative Tree (CAT) algorithm. Essentially, the algorithm first extracts and retains association classification rules from the training data which satisfy resource constraints, and then uses the rules to construct the final decision tree. The approach has advantages over the traditional top-down approach, first because only feasible classification rules are considered in the tree induction and, second, because their costs and resource use are known. In contrast, in the top-down approach, the information is not available for selecting splitting attributes. The experiment results show that the CAT algorithm significantly outperforms the top-down approach and adapts very well to available resources.
Read full abstract