-
Notifications
You must be signed in to change notification settings - Fork 0
Error Correcting Tournament (ect) multi class example
ECT stands for "Error Correcting Tournament".
The option --ect <K> where <K> is the number of distinct classes
directs vw to perform K multi-class (as opposed to binary)
classification. This methods differs from --oaa <K> only in
the algorithm used. The data input format is the same.
- Data-set labels must be in the natural number set {1 .. <K>}
- <K> is the maximum label value, and must be passed as an argument to
--ect
- See the paper http://arxiv.org/abs/0902.3176
Assume we have a 3-class classification problem. We label our 3 classes {1,2,3}
Our data set ect.dat may look like this
1 ex1| a
2 ex2| a b
3 ex3| c d e
2 ex4| b a
1 ex5| f g
This is essentially the same format as the non multi-class case (classification or regression) where each label must belong to one of the {1..<K>} classes, i.e. a natural number between 1 and <K>. You may add weights to the example and the features, use name-spaces, etc.
We train:
vw --ect 3 ect.dat -f ect.model
Which gives this progress output:
final_regressor = ect.model
Num weight bits = 18
learning rate = 0.5
initial_t = 0
power_t = 0.5
using no cache
Reading datafile = ect.dat
num sources = 1
average since example example current current current
loss last counter weight label predict features
0.000000 0.000000 1 1.0 1 1 2
0.500000 1.000000 2 2.0 2 1 3
0.500000 0.500000 4 4.0 2 2 3
finished run
number of examples per pass = 5
passes used = 1
weighted example sum = 5.000000
weighted label sum = 0.000000
average loss = 0.600000
total feature number = 15
Now we can predict, using the same data set as our test-set:
vw -t -i ect.model ect.dat -p ect.predict
Similar to what we do in vanilla classification or regression.
The resulting ect.predict file has contents:
2.000000 ex1
2.000000 ex2
3.000000 ex3
2.000000 ex4
1.000000 ex5
Which is as expected: 'ex5' belong to class 1, 'ex1', 'ex2' and 'ex4' belong to class 2, and 'ex3' belongs to class 3.