中文摘要：

作为拉普拉斯映射（LE）的线性近似而提出的局部保持投影（LPP）方法,同时具备流形学习方法和线性降维方法的优点,因而在数字识别、人脸识别等分类问题中受到广泛的关注。LPP构造近邻图主要有两种方法,其一,采用k近邻方式,其二,利用类别信息;采用k近邻方式没有利用已有的类别信息,而利用类别信息构造的方式没有充分考虑数据的局部几何关系。对此,提出一种基于改进邻域的局部保持投影方法（ILPP）,综合了k近邻方式和类别信息构造方式的优势,利用已知类别信息的同时又没有忽略流形的局部关系,明显地提高了LPP的分类效率。通过对比实验,表明了ILPP的有效性。

英文摘要：

Locality preserving projections（LPP）,presented as a linear approximation of Laplacian eigenmaps,are drawn wide attention in classification problems including digital recognition and face recognition because of its advantages in both manifold learning method and linear dimension reduction method.LPP generally uses two methods to construct the neighbour graph,the method of using k-nearest neighbour,and the method of using classification information.The method using k-nearest neighbour does not take advantage of existing classification information,and the method of using classification information does not fully consider local geometry information.In this regard,a locality preserving projections method based on improvement of neighbourhood（ILPP） is discussed in the paper,which integrates the advantages of both two,while utilising the known classification information,it does not ignore the local relationship of manifold as well.The new method significantly improves the efficiency of the classification of LPP.Comparative experiments demonstrate the validity of ILPP.