Minimal Cost Attribute Reduction through Backtracking

Test costs and misclassification costs are two most important types in cost-sensitive learning. In decision systems with both costs, there is a tradeoff between them while building a classifier. Generally, with more attributes selected and more information available, the test cost increases, and the misclassification cost decreases. We shall deliberately select an attribute subset such that the total cost is minimal. Existing decision tree approaches deal with this issue from a local perspective. They benefit from immediately available test results, therefore objects falling into different branches may experience different tests. In this paper, we consider the situation where tests have delayed results. Since we need to choose a test set for all objects, the attribute reduction problem is defined from a global perspective. We propose a backtrack algorithm with three pruning techniques to find a minimal cost reduct. Experimental results indicate that the pruning techniques are effective, and the algorithm is efficient on a medium sized dataset Mushroom.