CCA三支决策模型的边界域样本处理

摘要

构造性覆盖算法（constructive covering algorithm，CCA）三支决策模型在学习过程中根据样本分布特征，自动产生正域、负域和边界域。该模型为边界域样本处理问题提供了新的思路。重点讨论了基于CCA的三支决策的边界域样本处理问题。对边界域样本处理提出了两种决策方案：一种为处理全部的边界域样本，给出了距中心最近原则、距边界最近原则、万有引力原则3种方法；另一种为处理部分的边界域样本，即只对满足一定条件的边界域样本作进一步的划分，这样使不满足条件的边界域样本仍保留在边界域，提高了边界域样本处理的正确率。用十字交叉法在5组数据集上对这两种决策方案进行了对比，实验结果表明，处理部分边界域样本时正确率更高，效果更好。%The three-way decisions model based on constructive covering algorithm (CCA) produces automatically three regions:positive regions (POS), negative regions (NEG) and boundary regions (BND) according to the distri-bution characteristic of samples. The model provides a new idea to deal with samples in BND for the theory of three-way decisions. This paper proposes two methods to deal with samples in BND:The first method deals with all samples in BND, which gives three principles, nearest to the center principle, nearest to the boundary principle and gravity principle;The second one deals with part of samples in BND, i.e., the samples which satisfy some condi-tions are dealt with, and the rest samples are still in the BND. The second method promotes the accuracy. This paper compares the two methods on five datasets using 10-fold cross-validation. The experimental results show that the effect and accuracy of dealing with part of samples in BND are better than those of dealing with all samples in BND.