一种结合关联限制的最近邻分类策略

Approach for Combining Nearest Neighbor Classification with Pairwise Constraints

研究关联限制在最近邻分类中的应用,提出结合关联限制的最近邻分类算法PCNN.算法分成两个阶段:首先通过自学习过程,成对地添加施加关联限制的样本对;然后再进行一般的最近邻分类.引入最大半径和有效距离,并进一步给出自学习时样本对的评估方法,并且基于实验结果进行了分析.由于来自运输企业的行车数据能够较容易地施加关联限制,本算法比行车数据分类算法CIRP更为经济.对4个UCI数据集的分类结果也显示了算法的有效性.

The application of pairwise constraints in the Nearest Neighbor （NN） classification process is studied,and Pairwise Constrained NN（PCNN） algorithm is proposed.The PCNN algorithm consists of two steps,the first is a self-training step,during which pairs of constrained samples will be labeled simultaneously;and the second is a general NN classification process step.The max radius and effective distance is defined.Based upon them,the evaluation method on sample pairs for PCNN′s self-training process is proposed,and is further studied based upon test results.As the vehicle traveling records from transportation enterprise are readily labeled by pairwise constraints,they can be classified by PCNN more economically than CIRP,a classification algorithm specialized on the classification of vehicle traveling data.The classification of four UCI datasets also supports the effectivity of PCNN.