Abstract: This paper proposes a novel algorithm named K-NCA based on Neighborhood Component Analysis(NCA). It uses NCA to learn a Mahalanobis distance measure and reduces the dimension of the input dataset. The algorithm defines a class imbalance factor and introduces K Nearest Neighbor(KNN) to compute the test sample’s class-conditional probability estimation. The sample’s class label is decided by this probability. A text classifier is designed to accomplish the algorithm. Experimental results show that K-NCA algorithm can improve the accuracy of text classification.

Key words: Neighborhood Component Analysis(NCA), distance metric learning, dimension reduction, K Nearest Neighbor(KNN), text classification

摘要： 在近邻元分析(NCA)算法的基础上，提出K近邻元分析分类算法K-NCA。利用NCA算法完成对训练样本集的距离测度学习和降维，定义类偏斜因子，引入K近邻思想，得到测试样本的类条件概率估计，并通过该概率进行类别判定，实现文本分类器功能。实验结果表明，K-NCA算法的分类效果较好。

关键词: 近邻元分析, 距离测度学习, 降维, K近邻, 文本分类

CLC Number: 

• TP18