I have run gain ratio attributeSelection on weka and also run gain ratio attributeSelection from Java machine learning library
" https://sourceforge.net/p/java-ml/java-ml-code/ci/a25ddde7c3677da44e47a643f88e32e2c8bbc32f/tree/net/sf/javaml/featureselection/scoring/"
The results in term of attribute score and order is quiet different , is this related to diversity discretization method used in both ?
For example,
Gain ratio with discretization method A:
0.509 1671
0.432 765
0.425 249
0.396 625
0.393 493
0.369 1423
0.345 1772
0.344 267
0.344 245
0.335 1042
0.33 513
Gain ratio with discretization method B :
765 0.291676685
493 0.280632207
1411 0.272024568
1423 0.264074648
249 0.24425128
897 0.238272511
1967 0.205044156
1772 0.203703632
1843 0.201867619
245 0.201800878
Firstly let's understand gain ratio, gain ratio(for an attribute at a given split) = information gain / intrinsic information.
Gain ratio varies not only with the attribute but also the split at which it is being calculated. Feature/attribute selection algorithms can either by top-down (all to few features) induced or bottom-up (none to significant features) induced. So depending on the algorithm the gain ratio for an attribute calculated can vary significantly. Now if time permits, the best approach would be to use multiple feature/attribute selection algorithms, do a majority vote for ranking the best features, so that you won't miss significant attributes due to the greedy algorithms.
Thanks for valuable information.
But how about discretization in my question??
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running