Credibility Of LDA

[amritbhanu]

IDEA:

ACTUAL
          T1        T2         T3      .. .. . . .
Doc1
Doc2
Doc3

PREDICTED - Selected from Dominant topic from doc topic distribution.
          W1        W2         W3      .. .. . . .
Doc1
Doc2
Doc3

**According to literature, If a document is asked to belong to one of the dominant 
topic (hard assignment), the top words from the dominant topic should be in the 
actual document. If not:
 - then the probability of dominant topic is very less and there might be other topic which 
can be made dominant.
- or the top words are wrongly selected. The weights of words could be better to find 
the same dominant topic.**

Experiment:

• Once top n words are selected from each topic, now those topics are represented with those n words.
• A dominant topic is selected to represent a document, we call that as actual.
• we will check for each topic which are now represented with n words. We will find most 'm' words out of those 'n' in a document. Whichever topic will have the most 'm' words, according to this, now that document is represented with this topic.

We have now x no of documents. For eg x=4, k(no of topics)=3
for x=4, we have [D1,D2,D3,D4]
Actual=[1,1,2,0]
Predicted=[1,0,2,0]
The score is = 2/4=0.50

Results:

• Higher the better

  Conclusion:

• tuned with top 7 words is performing much better than untuned (default, k=10) top 7 words.
• tuned with top 7 words is performing better or same than untuned (default, k=10) top 10 words.
• With tuning we have better top 7 words defining that topic.

[timm]

plz clarify:

• was this with using lda as the terms for a subsequent use of SVM?
• the above results show 5 cases where tuend wa as good or better than other things. so why are you reporting this as a negative result?

[amritbhanu]

We have 2 tracks in lda now:

• for reporting stable conclusions. (related to model stability)
• another one for using LDA features into svm. (related to classification)

This one is related to the first track. We want to report stable topics generation and that only top 7 words are important after tuning rather than reporting 10 words with default.

I am reporting positive results.

[timm]

ur reporting positive results for...

1. for reporting stable conclusions.
2. another one for using LDA features into svm.

[amritbhanu]

just for the first right now.