J4 ›› 2014, Vol. 41 ›› Issue (3): 123-130.doi: 10.3969/j.issn.1001-2400.2014.03.018

• 研究论文 • 上一篇    下一篇



  1. (1. 西安电子科技大学 计算机学院,陕西 西安  710071;
    2. 陕西师范大学 远程教育学院,陕西 西安  710062)
  • 收稿日期:2013-03-13 出版日期:2014-06-20 发布日期:2014-07-10
  • 通讯作者: 李娟
  • 作者简介:李娟(1979-),女,讲师,西安电子科技大学博士研究生,E-mail: ally_2004@126.com.
  • 基金资助:


New nearest neighbor affinity similarity function based on separation and compactness between samples

LI Juan1,2;WANG Yuping1   

  1. (1. School of Computer Science and Technology, Xidian Univ., Xi'an  710071, China;
    2. School of Distance Education, Shaanxi Normal Univ., Xi'an  710062, China)
  • Received:2013-03-13 Online:2014-06-20 Published:2014-07-10
  • Contact: LI Juan



关键词: 机器学习, 近邻, 亲和相似度, 分散度, 紧密度


Traditional distance and similarity measurements did not take into account the influence of the individual sample on the whole sample set. To deal with this issue, a new similarity improvement strategy of k-nearest neighbor algorithm (KNN) is proposed in the paper. First, a new affinity distance function is introduced, which focuses on the separation and compactness between each individual sample and the whole sample set. Second, a new similarity function using this affinity distance function is proposed and taken as the similarity measure function in the KNN. Third, a theoretical analysis of and experiments on eighteen numerical UCI (University of California Irvine) datasets are made to compare the affinity similarity function proposed in this paper with classical distance or similarity functions through 5-fold partitioning cross-validations. Finally, classification results indicate that the proposed affinity similarity function is not only an effective similarity strategy for classification, but can reduce the classification time for large-scale data sets by combining efficient indexing algorithms.

Key words: machine learning, nearest neighbors, affinity similarity, separation, compactness