Course:CPSC532:StaRAI:2017:Zilun Peng

Course:CPSC532:StaRAI

Note: Tuffy results are different from what I reported before

Method	ml-60k ASE	ml-60k Log loss	ml-1m ASE	ml-1m Log Loss	Yelp ASE	Yelp Log Loss
Predict 0.5	0.25	1	?	?	?	?
training average	0.2159	0.9004	?	?	?	?
MLN/Rating wgts per item (cheating)	0.18699	0.7962	0.1347	0.6507	0.1867	0.7964
MLN LR/Rating (wgts per item+select #iter by cv)	0.1871	0.8089	0.1347	0.6057	0.1873	0.7923
Tuffy (model 1)	0.6499	inf*	0.3892	inf	0.2536	1.011

inf because Tuffy predicts a male to be 100% female and vice versa

MLN wgt per item

One rule for g(u)

One rule for each positive item Rpos(u,i) ^ g(u)

One rule for each negative item Rneg(u,i) ^ g(u)

Rpos(u,i) is true if user u gives a rating >=4 to item i

Rneg(u,i) is true if user u gives a rating <4 to item i

3000 iterations for 100k's results

3000 iterations for 1m's results

55000 iterations for yelp's results

Code:

https://github.com/zilunpeng/agg_exp/blob/master/mln_direct_LR_wgt_for_each_item.py

MLN LR/Rating (wgts per item+select #iter by cv)

select number of training iterations by cross validation

CV result (100k) #iterations = 6775

CV result (1m) #iterations = 2900

CV result (Yelp)#iterations = 64400

Code:

https://github.com/zilunpeng/agg_exp/blob/master/mln_LR_wgtsForEachItem_cv_100k.py https://github.com/zilunpeng/agg_exp/blob/master/mln_LR_wgtsForEachItem_cv_1m.py https://github.com/zilunpeng/agg_exp/blob/master/mln_LR_wgtsForEachItem_cv_yelp.py

Note: contents of those codes are the same except the parameters to find_min

Tuffy results

All Tuffy related files (codes, input files to Tuffy, etc) can be found at here: https://github.com/zilunpeng/agg_exp/tree/master/tuffy_exp

https://github.com/zilunpeng/agg_exp/blob/master/tuffy_exp/instructions.md contains commands that I used to learn weight and do inference on Tuffy