面向代价敏感的多标记不完备数据特征选择算法

Multi-label Feature Selection Algorithm with Incomplete Data Based on Cost Sensitivity

代价敏感下的特征选择是机器学习和数据挖掘领域的重要研究内容,目前基于代价敏感的特征选择研究主要是面向单标记的数据,由于在许多应用领域数据往往是多标记连续型数据,且在数据获取过程中由于技术或成本限制导致数据呈现出不完备性.为解决上述问题,提出了一种基于测试代价的多标记不完备数据特征选择算法.首先,算法利用粗糙集模型计算多标记不完备数据下的邻域粒度,并用均匀分布和正态分布两种分布函数计算每个特征的特征代价;然后,提出了一种基于测试代价的特征重要性计算方法,并在核特征的基础上,设计了启发式的特征选择算法;最后,通过在Mulan数据集上的实验结果进一步验证了算法的有效性和可行性.

The feature selection based on cost-sensitive is an important research issue in machine learning and data mining. At present,most feature selection work based on cost sensitivity deal with the single-label data. However,the data usually is multi-label and continuous in many applications,due to the technology or cost limitations during data collection,the data is incomplete. To alleviate this problem,a feature selection algorithm for multi-label incomplete data based on test cost is proposed in this paper. At first,the neighborhood granularities of the multi-label incomplete data are computed by the rough set model,and the test cost of each feature are generated according to the uniform distribution and normal distribution. Then,we calculate the core of the feature and put the core to the subset of the feature,at the moment,we redefine the feature criterion of significance degree based on the cost-sensitive is designed,and a heuristic algorithm of the feature selection is designed based on the core of the feature. Finally,the effectiveness and feasibility of the proposed algorithm is verified by the experimental results on the Mulan datasets.