自适应邻域图的流形学习方法

MANIFOLD LEARNING METHOD FOR ADAPTIVE NEIGHBOURHOOD GRAPH

针对目前流形学习方法的嵌入效果非常敏感于局部邻域的选取方式,提出一种自适应邻域图的非线性数据降维方法。该方法考虑数据点周围的点分布信息,自适应地寻找最近邻域大小。不同于传统的邻域选取方法,此方法根据样本点周围的疏密程度来动态地获得最近邻域数,且所得到的各个样本点的邻域数是不等的;将每个样本点与其最近邻点连接,构建自适应邻域图进行有效降维。在人工生成数据集和人脸数据上的仿真结果表明,提出的方法得到了良好的降维效果。

Embedded results of existing manifold learning methods are very sensitive to the selection of local neighbours.In light of this, we propose a non-linear data dimensionality reduction method for adaptive neighbourhood graph.It adaptively searches the nearest neighbour-hood size by considering data distribution information around each data point.Unlike traditional neighbourhood selection method,the pro-posed approach automatically derives the number of the nearest neighbours according to sparse or dense degree around each sample point.The derived number of the nearest neighbours for different sample points is unequal.Each sample point is connected with its nearest points.This constructs an adaptive neighbourhood graph,which can effectively reduce data dimensions.Results of simulation on the artificially generated data sets and face data show that the proposed method reaches a better dimensionality reduction effect.