Hello, i'm learning how to use VW and I do:
$ vowpal_wabbit/vowpalwabbit/vw.exe -d VWFood.txt \
--active_simulation --active_mellowness 100000
on this dataset: VWFood.txt.
It is based on eHarmony sample, slice 1, user #7757 compatibility generalization and downsampling. I parsed eHarmony data, trained an mlpy svm on #7757-containing pairs in the "labels" file, applied it to classify all users in the "data" file, then removed 99,9% of the negative class.
What i get is:
finished run
number of examples = 97
weighted example sum = 97
weighted label sum = -63
average loss = 0.227427
best constant = -0.649485
total feature number = 3216
total queries = 0
Which is wrong because 0 < total queries <= 97 should be the case.
I set active_mellowness to a large number to force quering all labels.
Moreover. In a realtime scenario (with --active_learning --daemon), it queries everything even with tiny active_mellowness (the opposite of what i'm describing above). But also when I double, triple etc. the dataset, VW keeps giving wrong predictions even for the examples it has seen (so no active and no learning).
What am I doing wrong? Anyone to teach me using Vowpal Wabbit efficiently in an AL setting?
Sorry for my english and lack of skill
asked
Jul 28 '13 at 12:54
Pasha M
1●1●1●1
