Multi Goal Tuning

[amritbhanu]

Experiment:

• Tuned for Raw score(Jaccard) and F2score
• With less than 30 Evaluations

  Conclusions:

  PROS:

• Dramatic Feature Reduction
• Better f2 score
• Now we can conclude as well as good prediction.
• With tuning, you find the right k, by automatic methods.

  CONS

• 10 times more runtime

  Results

• Now Topics can be seen as well with 7 words (https://github.com/ai-se/LDAClassification/blob/master/src/2016-10-11/Topics/Tuned/scifi.md)

Datasets	k=10 with w=10 (default) (raw score/f2score)	k=20(manual tweaks)	k=40	k=80
SE6	(0.22, 0.52)	(0.22, 0.47)	(0.11, 0.80)	(0.11, 0.84)
SE0	(0.22, 0.93)	(0.33, 0.94)	(0.11, 0.93)	(0.22, 0.91)
photo	(0.22, 0.52)	(0.11, 0.61)	(0.22, 0.41)	(0.11, 0.42)
scifi	(0.44, 0.63)	(0.33, 0.75)	(0.22, 0.80)	(0.11, 0.87)
rpg	(0.11, 0.47)	(0.11, 0.23)	(0.11, 0.21)	(0.11, 0.24)
cs	(0.22, 0.40)	(0.44, 0.24)	(0.11, 0.29)	(0.11, 0.58)
diy	(0.22, 0.48)	(0.11, 0.49)	(0.11, 0.60)	(0.11, 0.60)

[amritbhanu]

cdom selection