基于测地距离的GIDGC-KNN不平衡数据分类器

摘要

IDGC (imbalanced data gravitation classification)using Euclidean distance calculation in imbalanced data classification gravitational process ignores distribution traits of the data and class neighborhood correlation of the test data,leading to the problem of low accuracy.GIDGC-KNN unbalanced data classification model based on geodesic distance was proposed.On the ba-sis of IDGC gravitational amplification factor (amplified gravitation coefficient,AGC),the geodesic distance was combined with and KNN (K-nearest neighbor)algorithm to get implicit global data distribution and neighbor sample geometry class correlation. The model can not only adapt to the high-dimensional manifold with data structure,but also inherit the KNN neighbor sample class correlation advantages.Experiments on the KEEL datasets show that,compared with the IDGC algorithms,data-level al-gorithms,price-sensitive learning algorithms and integration methods algorithms,GIDGC-KNN unbalanced datasets have signifi-cant classification performance and better generalization ability in high data sets than in low-imbalanced data sets.%IDGC (imbalanced data gravitation classification)在不平衡数据分类中使用欧式距离计算引力时,未能考虑数据分布性状和待测数据近邻类相关性,存在准确率低的问题,为此提出一种基于测地距离的GIDGC-KNN不平衡数据分类器模型.在IDGC基础上引入放大引力系数 (amplified gravitation coefficient,AGC),结合测地距离和KNN (K-nearest neigh-bor)算法得到数据分布隐含的全局几何结构和近邻样本类相关性.该模型适应高维具有流形结构的数据,继承了KNN近邻样本类相关性的优点.KEEL数据集上的实验结果表明,与IDGC算法、数据层面算法、代价敏感学习算法和集成方法算法相比,GIDGC-KNN在高不平衡数据集上比在低不平衡数据集上有明显的分类性能优势和良好的泛化能力.