基于误差和可变代价的最优属性子集选择方法

Optimal Feature Subset Selection with Errors and Variable Costs

代价是现实数据的重要方面.数据的测试代价与数据的误差范围,即数据的粒度紧密相关,而误分类代价又跟测试代价有关,已有的属性选择方法往往忽视了这一点.为了处理这种情况,提出了一种基于误差范围和可变代价的最优属性子集选择方法.首先建立了该方法的理论框架,再设计了相应算法.在该方法中,测试代价和误分类代价根据不同的误差置信水平自适应地生成.再以最小化平均总代价为目标进行属性选择,从而得到最优的属性子集和误差置信水平.实验结果验证了所提方法的有效性.

Cost is important to data in real application. Test costs of data are sensitive to error ranges, namely, the granularity of data, while misclassification costs are also related to test costs. The existing feature selection methods often ignore it. To address this situation, an approach is proposed for selecting optimal feature subset with error ranges and variable costs. Firstly, the theoretical framework is established. Then the corresponding algorithms are designed. In the method, test costs and misclassification costs are adaptively computed according to the confidence level of measurement errors. The objective of feature selection is to minimize the average total cost. By the method, the optimal feature subset and the best confidence level of errors can be obtained. The experimental results manifest the effectiveness of the proposed approach.