基于局部相关维度的流形离群点检测算法

• 作 者: 黄添强 李凯 郭躬德
• 机 构: 福建师范大学数学与计算机科学学院计算机科学系,清华大学计算机科学与技术系,
• 出 处: 模式识别与人工智能
• 发表日期: 2011年05期
• 关键词: 数据挖掘 离群点检测 流形学习 局部相关维度

传统的离群点检测算法不适合检测流形离群点,目前专门针对流形离群点检测的算法报道较少.为此,基于实验观察的启示,提出用流形局部相关维度检测流形离群点的算法.首先探讨内在维度的性质,并基于实验观察提出用流形局部相关维度来度量流形离群点,然后证明流形局部相关维度可表征数据样本离群的性质,最后基于此性质提出流形离群点检测算法.在人工数据与真实数据上的实验表明本算法可检测流形离群点,且本算法比最近报道的流形除噪算法具有更优的性能.

The traditional outlier detection algorithm is not suitable for detecting manifold outliers, so there are few reports on outlier detection algorithms.For this reason, based on the experimental observations, we propose to detect the manifolds by using the local correlation dimension of manifolds Outlier algorithm.Firstly, we discuss the properties of inner dimensions and propose to measure the outlier of manifolds by using the local correlation dimension of manifolds based on experimental observation.Finally, we prove that the local correlation dimension of manifolds can characterize the data sample outliers. Finally, This property proposes a manifold outlier detection algorithm.Experiments on artificial data and real data show that the proposed algorithm can detect manifold outliers and has better performance than the recently reported manifold de-noising algorithm.