How does weka handle overfitting in the information gain calculation? -
How does weka handle overfitting in the information gain calculation? -
i messing around info gain attribute evaluator of weka when stumbled on inconsistency when handling overfitting attributes. in relation utilize 3 numeric attributes , class attribute divides between failing , right info sets. class attribute depends on attr3 while other 2 attributes have random values. arff file:
@relation test @attribute attr1 numeric @attribute attr2 numeric @attribute attr3 numeric @attribute class {failing,correct} @data 1, 12.714813, 5000, failing 26, -10000, 0, right 10, 9.547521, 5000, failing -10000, 3.699694, 0, right 38, 6.1541, 5000, failing after executed info gain attribute evaluator default settings, output ranking:
=== attribute selection on input info === search method: attribute ranking. attribute evaluator (supervised, class (nominal): 4 class): info gain ranking filter ranked attributes: 0.971 3 attr3 0.971 2 attr2 0 1 attr1 selected attributes: 3,2,1 : 3 now, attribute 3 rightfully has info gain of 0.971 equals entropy of relation. evaluating other 2 attributes info gain not create sense contain highly mutual information. not understand, why 1 overfitting attribute has maximum info gain while other 1 has minimum info gain. how info gain attribute evaluator handle overfitting attributes?
weka information-theory
Comments
Post a Comment