一种改进的局部均值伪近邻算法

Improved Local Mean-Based Pseudo Nearest Neighbor Algorithm

针对基于局部均值的伪近邻分类算法(LMPNN)易受近邻参数k和噪声点影响的问题,提出了一种改进的局部均值伪近邻分类算法(IPLMPNN)。利用双层搜索规则确定待测样本的最近邻,提高近邻集的选择质量;为了克服主观赋权法的不利影响,并且加强每个局部均值向量对分类的作用,引入注意力机制计算距离加权系数;使用改进的调和平均距离计算待测样本与局部均值向量之间的加权多调和平均距离,由此查找伪近邻点对待测样本进行分类。利用UCI和KEEL中的多个数据集对IPLMPNN算法进行仿真实验,并与8种相关算法进行比较。实验结果表明,IPLMPNN算法取得了令人满意的分类结果。

Since the local mean-based pseudo nearest neighbor classification algorithm(LMPNN)is easily influenced by the parameter k of neighbors and noise points,an improved local mean-based pseudo nearest neighbor classification algo-rithm(IPLMPNN)is proposed in this paper.Firstly,the nearest neighbors of the test sample are selected by using the two-layer search rule to improve the quality of selecting neighbors.Secondly,in order to overcome the adverse influence of the subjective weighting method and strengthen the effect of each local mean vector for classification,the attention mecha-nism is introduced to calculate the distance weighting coefficients.Finally,an improved harmonic mean distance is provided to calculate the weighted multi-harmonic mean distance between a test sample and a local mean vector.Furthermore,the pseudo nearest neighbor points are found by the weighted multi-harmonic mean distances to classify test samples.The pro-posed algorithm is assessed and compared with relative algorithms for multiple data sets from UCI and KEEL.The experi-mental results show that IPLMPNN obtains satisfactory classification results.