mi5victim / greekmedia

Kaggle Competion - Multi-label prediction

#Multi-label classification of print media article to topics.#

6/9/2014 2:13:52 PM

https://www.kaggle.com/c/wise-2014

There are 203 class labels, and each instance can have one or more labels. We convert this problem to a binary classification problem so that vowpal wabbit can handle it.

The data in libsvm format, we use scripts/feature_creation.py to convert them to vw format. For every instance, say

103 1123:0.003061 1967:0.250931 3039:0.074709 20411:0.025801 24432:0.229228 38215:0.081586 41700:0.139233 46004:0.007150 54301:0.074447 .......

We create 203 vw entries as follows

....
-1 |LABEL_102 1123:0.003061 1967:0.250931 3039:0.074709 20411:0.025801 24432:0.229228 38215:0.081586 41700:0.139233 46004:0.007150 54301:0.074447 .......NO_WORDS:17

+1 |LABEL_103 1123:0.003061 1967:0.250931 3039:0.074709 20411:0.025801 24432:0.229228 38215:0.081586 41700:0.139233 46004:0.007150 54301:0.074447 .......NO_WORDS:17

As seen above we have added a new feature NO_WORDS, to count the number of words.

Using vw we train it as follows

vw --loss_function hinge  -d data/wise2014-train.vw --binary  -f models/model-09-13-18.bin

number of examples per pass = 13165768
passes used = 1
weighted example sum = 1.31658e+07
weighted label sum = -1.29777e+07
average loss = 0.00458097
best constant = -0.985716
total feature number = 3519146694

For feature prediction, we run the model in daemon mode

vw -i models/model-09-13-18.bin --daemon --quiet -t --port 26543

scripts/prediction.py is used to predict the test set.

This puts us in 16th place.