基于最小联合互信息亏损的最优特征选择算法  

Minimum Joint Mutual Information Loss-based Optimal Feature Selection Algorithm

作　　者：张逸石[1] 陈传波[1,2] 

机构地区：[1]华中科技大学软件学院,武汉430074 [2]华中科技大学计算机科学与技术学院,武汉430074

出　　处：《计算机科学》2011年第12期200-205,共6页Computer Science

基　　金：国家自然科学基金项目(60973085)资助

摘　　要：提出了一种基于最小联合互信息亏损的最优特征选择算法。该算法首先通过一种动态渐增策略搜索一个特征全集的无差异特征子集,并基于最小条件互信息原则在保证每一步中联合互信息量亏损都最小的情况下筛选其中的冗余特征,从而得到一个近似最优特征子集。针对现有基于条件互信息的条件独立性测试方法在高维特征域上所面临的效率瓶颈问题,给出了一种用于估计条件互信息的快速实现方法,并将其用于所提算法的实现。分类实验结果表明,所提算法优于经典的特征选择算法。此外,执行效率实验结果表明,所提条件互信息的快速实现方法在执行效率上有着显著的优势。In this paper, a minimum joint mutual information loss-based optimal feature selection algorithm was pro- posed, which firstly finds a non-discriminate feature subset of the original set via a dynamic incremental searching strate- gy, and then eliminates false positives by keeping minimum joint mutual information toss with class in each iteration u sing a minimal conditional mutual information criterion, in such a way as to obtain an approximate optimal feature sub set. Furthermore, for the computationally intractable problem arising in high dimensional feature space that characterizes the existing method of conditional independence test with conditional mutual information, a fast implementation of con ditional mutual information estimation was introduced and used to implement the proposed algorithm. Experimental re- sults for the classification task show that the proposed algorithm performs better than the representative feature selec- tion algorithms. Experimental results for the execution task show that the proposed implementation of conditional mutual information estimation has a considerable advantage.

关 键 词：特征选择 条件互信息 最小联合互信息亏损 快速实现 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]