H-LDA (Blei C++) 关键代码分析

1. Sampling levelallocations  (all prob. is calculated and represented by log)

Parameters:

GEM(m,π):

m:  gem_mean (mean of the stick leagth) ;

π:  gem_scale (variance of the stick leagth) .

 

Larger m, more likelyassign more words from each doc to higher levels of abstraction.

Larger π, wordallocations will not likely deviate from such a setting.

 

In the stick-breakingprocess:

  Θ ~ GEM(m,π);

  V_i ~ Beta(mπ, (1-m)π).

2. Sampling Paths