suehara-tsuyoshi / pattern-recognition-adv. Goto Github PK

Please use the following command to run algorithms after moving to ML_est or Bayes_est directory.

python em_gmm.py x.csv z.csv params.dat

python vb_gmm.py x.csv z.csv params.dat

I describe contents of the outputs z.csv and params.dat here.

This file stores the posterior probabilities of latent variables. Each row corresponds to observed data and each column corresponds to a class.

This file stores (variational) parameters obtained through each algorithm. Here, I describe the final value of each parameter, paired with its parameter name.

What is the appropriate number of classes K ? I ran the EM algorithm with different values of this parameter.

As the above figure shows, the log-likelihood is maximized to　almost the same value when I set the number of classes to equal to or more than 4. Therefore, K=4 is the optimal value for this dataset.

In fact, it can be reasonable from the figure for clustering using EM algorithm with K=4.