Abstract: Curse of dimensionality often comes out when learning from high dimensional data.Locality Preserving Projection（LPP） can preserve the local geometry of the manifold of data and solves this problem well.However，the similarity utilized in LPP is sensitive to outliers and noise，thus an Enhanced Locality Preserving Projection（ELPP） method is proposed with the robust path based similarity.Experiments on high dimensional manifold data prove the effectiveness of this method.

Key words: curse of dimensionality, feature extraction, similarity metric, robustness

摘要： 维数灾难是机器学习算法在高维数据上学习经常遇到的难题，基于局部保持的投影方法（Locality Preserving Projection，LPP），可以很好地解决维数灾难难题。然而传统LPP的相似性度量方法对噪音敏感，为此利用鲁棒路径相似的度量方法，提出一种增强的局部保持投影方法。在高维流形数据上的降维实验证实了该方法对噪声和离群点的有效性。

关键词: 维数灾难, 特征抽取, 相似性度量, 鲁棒性