一种选择标注分层流形学习算法

A Selecting Landmark Hierarchical Manifold Learning Algorithm

流形数据的查询需要使用流形的嵌入表示,因此查询流形数据需要访问大量的样本数据.提出一种选择标注分层流形学习算法,选择出的标注点集用来帮助查找流形数据.首先采用自适应近邻算法求出每个数据的最优近邻,然后构造测地线矩阵,最后逐步迭代随机选择标注点,求出每个标注点的极大单元子集,直到流形数据集变成空集,形成初始标注点集.此外,还要优化标注点集.实验结果证明所选择的标注点集保持流形的拓扑特性,可有效帮助查询流形数据.

The manifold data query needs the manifold embedded representation. Thus it often involves accessing considerable volume of data. An approach of hierarchical manifold learning algorithm based on selecting landmark points from the given samples is proposed for representing data on manifold. The landmarks set can help locate the novel points on the data manifold. Firstly, an adaptive nearest neighbor＇s method is employed to extract the nearest neighborhood of each data. Then the geodesic matrix is constructed. Finally, a landmark point is randomly selected in landmark point set, and its maximum cell is found till the manifold set is empty and the rough landmark point set is formed. In addition, the landpoint set is optimized. The experimental results prove that the proposed method preserves the topological features of manifold, and it helps inquire the manifold data efficiently.