LLE(Locally Linear Embedding)算法

Core idea

LLE is inherently a non-linear dimensionality reduction strategy

即局部线性嵌入算法。该算法是针对非线性信号特征矢量维数的优化方法,这种维数优化并不是仅仅在数量上简单的约简,而是在保持原始数据性质不变的情况下,将高维空间的信号映射到低维空间上,即特征值的二次提取。

Charateristics: neighborhood-preserving

Steps

  • Select neighbors
  • Reconstruct with linear weights
  • Map to embedded coordinates

Basic formulas

How to translate this problem to eigenvalue solution

Use a Lagrange multiplier

More details can be found on https://segmentfault.com/a/1190000016491406

Graph Embedding

  • 寻找neighborhood:直接用Graph的邻接结构表示neighborhood
  • 计算linear weights:直接用邻接矩阵W
  • 生成embedding:计算矩阵M特征值,当节点数为n,embedding为q维时,取[n-q, n-1]的特征向量为embedding结果

Code

simple_code:

another version:

 

留下评论

您的电子邮箱地址不会被公开。 必填项已用 * 标注