A genetic algorithm to the minimal test cost reduct problem

Cost-sensitive learning exists in many data mining and machine learning applications. It considers various types of costs, such as test costs and misclassification costs. The test-cost-sensitive attribute reduction problem attracts our interest. It aims at finding a minimal cost test set, which preserves whole information of the decision system. An existing heuristic algorithm is proposed to address the problem. However, the results are unsatisfactory in larger datasets. Since genetic algorithms provide robust search in complex spaces and work well on combinatorial problems. In this paper, we propose a new approach based on the genetic algorithm to produce better results. In the algorithm, the fitness function is constructed based on the number of selected conditional attributes, the positive region, test costs and a user-specified non-positive exponent λ. A number of reducts are produced with different λ settings. Then the best reduct is selected as the suboptimal reduct. We compare the performance of the new approach with the existing one through experiments in four UCI (University of California-Irvine) datasets. Results show that the new approach generally produces better results and is more appropriate for medium-sized datasets than the existing one. The new algorithm can be further combined with the existing one to produce even better results.