EP vs. VB plots

The vocabulary has 3 words. The probability of word 1 vs. word 2 is plotted (the probability of word 3 is redundant). Dataset always has 10 docs of length 100, sampled from the model.

True aspects equal

True aspects slightly spread out

More spread out

Two aspects (fewer aspects than words)

Doc len 10,000

100 docs of len 10,000

100 docs of len 100

Test problem

3 aspects over 5 words, alpha = 1.

Initial aspects:

  1.8705533e-001  1.9462531e-001  1.7976830e-001  2.2720505e-001  2.1134602e-001
  2.0031929e-001  1.8673649e-001  2.1593019e-001  1.8391741e-001  2.1309662e-001
  1.8332725e-001  1.8120753e-001  2.2369119e-001  2.1850893e-001 1.9326510e-001

10 docs of 100 words each:

A A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B
B B B B B C C C C C C C C C C C C C C C C C D D D D D D D D D D D D D D D D
D D D D E E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B B
B B B B B B B B C C C C C C C C C C C C C C C C C D D D D D D D D D D D D D
D D E E E E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B B B C C C
C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C C D D D D D D D
D D D D D D D D D E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B B B B B B B B
B B B B B C C C C C C C C C C C C C C C C C C C C C C C D D D D D D D D D D
D D D D D D D D D D D E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A B B B B B B B B B B B B C C C C C C C C C
C C C C C C C C C D D D D D D D D D D D D D D D D D D D D D D D D D D D D D
D E E E E E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B C
C C C C C C C C C C C C C C C C C C D D D D D D D D D D D D D D D D E E E E
E E E E E E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B
B B B B B B B B C C C C C C C C C C C C C C C C D D D D D D D D D D D D D D
D D D D E E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B
B B B B C C C C C C C C C C C C C C C C C C D D D D D D D D D D D D D D D D
D D D D D E E E E E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B B B B
B B B B C C C C C C C C C C C C C C C C C C C C C C C D D D D D D D D D D D
D D D D D D D D D E E E E E E E E E E E E E E E 
A A A A A A A A A A A A A A A A A A A A A A A A B B B B B B B B B B B B B B
B B B B B B B C C C C C C C C C C C C C C C C D D D D D D D D D D D D D D D
D D D E E E E E E E E E E E E E E E E E E E E E

EP

Aspect 1 parameters versus iteration:

Converged result:

    0.1960    0.2300    0.1575
    0.1339    0.2517    0.2443
    0.1749    0.1446    0.2910
    0.2549    0.1540    0.1822
    0.2403    0.2197    0.1251

VB

Aspect 1 parameters versus iteration:

Converged result:

    0.2619    0.3000    0.0175
    0.0000    0.5389    0.1050
    0.0113    0.0086    0.6045
    0.3061    0.0041    0.2712
    0.4207    0.1483    0.0017

Tom Minka

Last modified: Mon Aug 22 18:00:48 GMT 2005