Dimension Reduction Technique
Embed higher-dimensional data on representative lower-dimensional manifold.
In this project we sought to find lower dimensional structure presented in higher-dimensional data through various nonlinear dimension reduction techniques. Diffusion map adopts a probabilistic approach to dimension reduction by performing random walk on the high dimensional feature space. I investigated its effectiveness in dimension reduction on simulated data and real image data. The left figure shows diffusion map reducing high dimensional image data to a 1D representation, which can be interpreted as the rotation angle of the images.
