关键词: 不平衡数据集, 过采样, 自然最近邻, 聚类

Abstract:

Aiming at the problem of introducing noise points and synthesizing overlapping samples in existing oversampling methods, this paper proposes an oversampling method based on natural nearest neighbors. The proposed method firstly determines the natural nearest neighbor for minority samples. Each sample’s number of nearest neighbors is generated by adaptive calculation in the algorithm, which reflects the density of distribution. After cluster analysis for minority samples based on relations of natural neighbor, this method generates new samples using core points in dense area and non-core points in sparse area from the same cluster. The comparison experiments on a two-dimensional synthesis dataset and UCI datasets verify the feasibility and effectiveness of this method and improve the classification accuracy of unbalanced data.

Key words: imbalanced data set, over sampling;natural nearest neighbor, clustering