基于离群点检测的鲁棒局部切空间排列方法

Robust Local Tangent Space Alignment Based on Outlier Detection

研究局部切空间排列方法(LTSA)对离群点的敏感性,提出一种基于离群点检测的鲁棒局部切空间排列方法(RLTSA).该方法用样本点到切空间的投影距离检测离群点.在构造样本点局部邻域时,RLTSA尽可能排除离群点,以构造稳定的局部邻域,而对离群点,RLTSA把它们投影到更高维的切空间,以减少离群点的投影距离.模拟实验和实际例子说明,新方法能提高局部切空间排列方法处理离群样本点的能力.

The paper focuses on the sensitivity of local tangent space alignment （LTSA） to outliers, and presents a robust local tangent space alignment （RLTSA） based on outlier detection. RLTSA detects the outliers by the projection distances of the sample points onto their tangent spaces. When constructing the neighborhoods of the sample points, RLTSA removes the outliers to construct stably local neighborhood. For the outliers, RLTSA projects them onto higher dimensional tangent spaces to reduce their projection distances. Simulation and real examples show that the new approach can improve the ability of LTSA to deal with outliers.