2011_08_08

Post date: Aug 9, 2011 6:57:21 AM

There are many things to take care of:

  • Upload the 2D covariance matrix plot. fn_calculateEllipse, fn_plot_covariance_2D, plotEllipse2d [done]
  • Learn more about function gmmVBEM4 which uses MATLAB's gmm object. The function is the core of VBGMM clustering. [done]
  • Learn more about the function fn_VBGMM_clustering. See why the Sigma calculated from the code is very big??? [done]
    • I found the answer! That's because I forgot to inverse W matrix. Ideally, the Sigma would be inv(W(:,:,i))/(v(i)-D-1), but I mistakenly forgot the inverse, i.e., W(:,:,i)/(v(i)-D-1), which is wrong.
  • gmmVBEMdemo3 is the main code where everything is demo here. [done]

After VBGMM is done, I will implement fast DT using VBGMM

  1. Use VBGMM to cluster the data, and the number of clusters will be determined automatically with its corresponding mean and covariance matrix.
  2. We can make the optimal structure out of the VBGMM result
  3. Add a few more empty parent nodes on the root

Note that in this version we will assume correlated features, so we don't have to slice the feature anymore. Also, the number of layer is 2 not 3, so this should be simple and fast!

The deadline for my thesis would be Sep 1!!! So I have to keep writing.