Implementation of linear svm with l2-penalty and hinge loss, without interception.
Results of the perforamce on text classification tasks. The Web Track from TREC is used, to compare against the sklearn.svm.LinearSVM(C=10000, class_weight=None, dual=True). The documents are spitted, and 60% documents are used as training data.
The reported gain/loss is based on average F1 score from binary classification with GD for 50 queries with 300 iterations.
| Gain (+) /Loss (-)
------------------------------------------------------------------------------|-------- TREC 2011 Web Track: Topics 101-150 | -7.1% TREC 2012 Web Track: Topics 151-200 | -5.2% TREC 2013 Web Track: Topics 201-251 | -1.4% TREC 2014 Web Track: Topics 251-300 | +0.7%