基于遗传算法的多扰动多K近邻分类器系统

为改善维数灾难对K近邻分类器的影响，提出一种基于遗传算法(GA)的多扰动的K近邻融合算法，简称GA-MKNNC算法。目标扰动将所识别的问题划分成多个子分类问题进行单独识别。针对不同子分类问题，数据扰动选取相关的数据，特征扰动确定相关的特征，参数扰动明确相关参数值。数据扰动由Bagging算法确定。特征扰动和参数扰动通过GA学习得到。多个子分类问题的决策通过最大融合得到最终决策。实验结果表明，该算法的性能优于K近邻分类器及多数融合算法，且选用的子分类器数目少于FASBIR算法。

Multiple K Nearest Neighbor Classifiers System Through Multimodal Perturbation Based on Genetic Algorithm

In order to relax the curse of dimensionality, a new algorithm for K Nearest Neighbor Classifiers(KNNC) fusion through multimodal perturbation based on Genetic Algorithm(GA) is proposed, called GA-MKNNC. The unsolved problem is decomposed into multi sub-problems by target perturbation, and each sub-problem is independently classified. In the process of classification, one sub-problem is classified by the corresponding data, feature and parameters, which are called data perturbation, feature perturbation and parameter perturbation. Data perturbation selects data by bagging. Feature perturbation and parameter perturbation are learned by GA. The decisions from sub-problem classifiers are combined by maximum rule to get the final decision. Experimental results show GA-MKNNC is better than KNNC and most multiple classifier systems. In addition, the number of component classifiers in GA-MKNNC is less than that in FASBIR.