Cooperative Learning to Minimal Time Cost Attribute Reduction Through Ant Colony Optimization

摘要

One of the central problems in cost-sensitive learning is minimal time cost attribute reduction. Considering the problem, an artificial bee colony algorithm has been designed in recent research. However, in many cases, the performance of the algorithm is not satisfactory. In this paper, we design an ant colony optimization algorithm with the addition, deletion, and filtration stages to deal with this problem. A complete graph is adopted as a model with each vertex corresponding to an attribute and weight of each edge to pheromone. In the addition stage, each ant travels probabilistically until the positive region constraint is met. In the deletion stage, each ant deletes redundant attributes. Two strategies, called the centralized and distributed strategies, are developed. In the filtration stage, the ant with minimal time cost is selected to construct the optimal reduct. Experimental results on four UCI datasets show that our algorithm outperforms the existing one.