基于局部性正则化推广误差界的特征选择算法

Feature Selection Based on Locality Regularized Generalization Error Bound

特征选择是当前模式识别领域的研究热点.滤波方法和封装方法是特征选择算法中评价特征子集的两种主要策略,但均不能保证其后所设计的分类器的推广性能.针对以上两种策略的不足,首先引入基于样本流形结构的局部性正则化推广误差界.并在此基础上,以局部性正则化推广误差界为评价函数,以局部性正则化分类方法为目标分类器,提出一种混合滤波-封装型特征选择算法.该算法既保持了较高的计算效率,又保证了目标分类器良好的推广性.实验结果表明,新算法具有比对比算法更优的分类性能.

Feature selection is a hot topic in current pattern recognition. Filter and wrapper approaches are two of the most important policies to evaluate feature subsets in feature selection algorithms. However, they both can not guarantee the generalization performance of the following designed classifier. To solve these problems in the two approaches, a locality regularized generalization error bound is firstly introduced which embeds the manifold structure information hidden in the input samples. Furthermore, a hybrid filter- wrapper feature selection algorithm is proposed, which uses the locality regularized generalization error bound as the evaluation function as well as the locality regularization method as the classifier. As a result, the proposed algorithm can not only keep high computational efficiency, but also guarantee the good generalization performance of the following classifier. Experimental results validate the superiority of the algorithm.