Cost-Sensitive Feature Selection Based on Adaptive Hunting Optimization

For machine learning tasks, the explosive growth of data makes feature selection be an essential preprocessing step, as it improves generalization capability, reduces run-time and model's complexity. Traditional feature selection methods select the informative subset to facilitate the classification accuracy. However, in real applications, the cost of collecting the features shall be taken into account. This paper proposes a new cost-sensitive feature selection method based on an adaptive optimization framework. Unlike most of the existing methods simply adding or deleting features one by one, the proposed method uses an adaptive swarm intelligence algorithm to search the optimal subset. This algorithm achieves a more reasonable balance between the exploration and exploitation utilizing a cosine congestion factor, and is employed in cost-sensitive feature selection problem. Both of test cost and misclassification cost are considered, then a new and more reasonable comprehensive evaluation criteria is proposed. Experimental results on benchmark data sets demonstrate the effectiveness of the proposed methods.