一种新的证据K-NN数据分类算法

A New Evidential K-Nearest Neighbors Data Classification Method

K近邻分类算法已被广泛应用于模式识别中。为了有效处理识别问题中的不确定信息并提高数据分类精度,提出了一种新的证据K-NN(NEK-NN)分类算法。首先从总的训练集中随机重复采样来构造多个训练样本子集。在每个训练子集中,利用目标数据与其各个近邻的距离分别构造基本置信指派,并根据K个近邻数据在每个类别中的数目来对构造的置信指派进行加权。然后,利用DS规则对加权证据融合。根据每个训练子集下融合结果的算术平均值来判断目标的类别属性。通过模拟数据集和真实数据集的实验,将NEK-NN算法与其他几种常见的方法做了对比分析,结果表明NEK-NN算法能够有效地提高分类的精度。

The K-Nearest Neighbor （K-NN） rule has been widely used in the pattern recognition field. In order to effectively deal with the uncertain information and to improve the accuracy of classification, a new evidential K-Nearest Neighbors （NEK-NN） data classification method is proposed. Several training subsets are resampled from the whole training set. In each subset, the basic belief assignments （bba＇s） are determined using the distance between the object and its K Nearest Neighbors, and then the K bba＇s are discounted according to the number of the K Nearest Neighbors in each class. Finally the discounted bba＇s are combined using DS rule, and the mean of these combination results in each training subset is used for the classification of the object. Several experiments are given to test effectiveness of NEK-NN with respect to some other methods. The results indicate that NEK-NN can effectively improve the classification accuracy.