Feature selection is a fundamental preprocessing step for many machine learning and pattern recognition systems. Notably, some mutual-information-based and correlation-based feature selection problems can be formulated as fractional programs with a single ratio of polynomial 0–1 functions. In this paper, we study approaches that ensure globally optimal solutions for these feature selection problems. We conduct computational experiments with several real datasets and report encouraging results. The considered solution methods perform well for medium- and reasonably large-sized datasets, where the existing mixed-integer linear programs from the literature fail.

特征选择是许多机器学习和模式识别系统的基本预处理步骤。值得注意的是，一些基于互信息和基于相关的特征选择问题可以表述为具有多项式0-1函数的单个比率的分数程序。在本文中，我们研究了确保针对这些特征选择问题的全局最优解决方案的方法。我们使用几个真实的数据集进行了计算实验，并报告了令人鼓舞的结果。对于中等规模和相当大的数据集，考虑到的解决方法效果很好，而文献中现有的混合整数线性程序均无法使用。