CAJ | 학술논문

Citation-kNN算法对传统的kNN算法进行了改进,使其可以应用于多示例学习问题,但其0-1决策方式具有一定的局限性,没有充分考虑样本的分布情况.为解决该问题,该文提出局部加权的 Citation-kNN 算法,综合考虑样本的分布情况,提出基于样本距离加权、基于样本离散度加权的方法,并对各种组合情况进行了实验.在标准数据集MUSK和乳腺超声图像数据库上的实验结果表明,该文提出的方法与Citation-kNN相比,性能有明显提高,并具有良好的适应性.
Citation-kNN산법대전통적kNN산법진행료개진,사기가이응용우다시례학습문제,단기0-1결책방식구유일정적국한성,몰유충분고필양본적분포정황.위해결해문제,해문제출국부가권적 Citation-kNN 산법,종합고필양본적분포정황,제출기우양본거리가권、기우양본리산도가권적방법,병대각충조합정황진행료실험.재표준수거집MUSK화유선초성도상수거고상적실험결과표명,해문제출적방법여Citation-kNN상비,성능유명현제고,병구유량호적괄응성.
The Citation-kNN algorithm improves traditional kNN algorithm and can be applied to solve multi-instance learning issue. But its 0-1 decision strategy has some limitations. To overcome this issue, the locally-weighted Citation-kNN algorithm is presented in this paper. Considering distribution of the samples, the distance-based weighted method and the scatter-based weighted method are proposed. And their combinations are discussed. The method is applied to the standard database MUSK and the breast ultrasound image database. The results confirm that the method has higher accuracy comparing with that by using Citation-kNN algorithm.