基于混合Hausdorff距离的多示例分类问题

The Problem of Multiple-instance Classification Based on the Integration Hausdorff Distance

在机器学习领域中,多示例学习是一个重要的研究方向,其显著特点是正包中示例的类别标记具有模糊性。基于不同Hausdorff距离的CKNN分类器在多示例学习中应用较为广泛。经分析可发现,最小和最大Hausdorff距离均有其各自的缺陷,但两者的缺陷具有一定的互补性。针对如何弥补单一Hausdorff距离缺陷的问题,使用AdaBoost算法思想,把基于最小和最大Hausdorff距离的CKNN分类器进行组合,以减少使用单一Hausdorff距离对实验结果造成的影响。通过比较在不同数据集上的实验结果,可知此方法在一定程度上降低了测试误差,降低幅度最大为0.110 0。

The multiple-instance learning has received growing attention in the machine learning research field. The typical feature of multiple-instance learning is that the label of positive bag^s instance is fuzzy. The CKNN clas-sifiers are commonly used to research the multiple-instance problem base on different Hausdorff distance. Although the minimal Hausdorff （ minH） and maximal Hausdorff （ maxH） distance have their own drawbacks, to some ex-tent, they can make up for each other. AdaBoost algorithm was used to combine the CKNN classifiers which are based on the different Hausdorff distances. The purpose of the combination is recovered the defects caused by any single distance. Experiment show that the combination is effective and can reduce the test error by 0. 110 0 at most.