Results - Githubissues

ai-se / Pits_lda

IST journal 2017: Tuning LDA

https://github.com/amritbhanu/LDADE-package

4 stars 4 forks source link

Results #18

Closed amritbhanu closed 8 years ago

amritbhanu commented 8 years ago

How to read graphs?

These graphs are (tuning - untuning) results
If y value 0, tuned and untuned results are same.
If y >0, tuning improved by that much y margin
if y <0, tuning didn't help and made it worse.
Results

F = 0.3, CR = 0.7, Pop = 10

file

F = 0.7, CR = 0.3, Pop = 10

file

F = 0.3, CR = 0.7, Pop = 30

file

F = 0.7, CR = 0.3, Pop = 30

file

Conclusion

Tuning helped for sure in most number of cases.
Termination is based on no. of iterations, they can be increased to get better results for less no of terms overlap or they can be reduced to make it faster.
On HPC, it took like 4 hours to generate each graph.

amritbhanu commented 8 years ago

@timm Prof. here are some results.

[ ] Conclusion no 2 point, need some clarification.
[ ] For SO, results are still running, hoping to get them by tomorrow.

amritbhanu commented 8 years ago

Results - Wikipedia

1% of full 2gb wikipedia (preprocessed and top 5% tfidf)
Different lines are variations with F, CR and pop size.

file

Points to ponder

Untuned results may not be accurate, as we did 1 run to find the scores. The shuffled row order input may produce better scores and % improvement can be less. This problem is with all the previous results and all the datasets. This can be solved by selecting the mean score / the maximum score after doing multiple runs.