Results on R2001(Topic part) data (101 categories in training set,103 categories in test set,excluding "None" category)
3.0 kNN(standard)
Result tuned for microf1 and Result tuned for macrof1
Result
tuned for microf0.5
and Result
tuned for macrof0.5
The setting is: k=100, fs=8000,
fbr=0.1(for macro avg. f0.5) and 0.5(for micro avg. f0.5)
3.2 Rocchio on 2001t
The
result tuned for microf1 and the
result tuned for macrof1 (I also tried rcut and the
result is much worse)
The
result tuned for both micro and macro f0.5
The
graph tuning feature selection number (5000 for both micro and
macro avg. performance)
The
graph tuning fbr score(0.3 for micro avg. performance and 0.2 for
macro avg. performance, generally, 0.2 is OK)
The
graph tuning pmax (3000 for both micro and macro avg.
performance)
The
graph tuning beta(-1 for both micro and macro avg. performance)
3.3 NB(rainbow)
The
result tuned for microf1 and the
result tuned for macrof1
(for micro avg. result, all the
features are used. fbr=0.3. For macro avg. result, 3000
top features are used. fbr=0.)
3.4 SVM on 2001t
Result
tuned for microf1
Result
tuned for macrof1
Result
tuned for macrof0.5
Conclusion:
