This is indeed a very interesting question. Here are my cents:information gain vs gini index
Given how both values are calculated (see e.g. here
), the difference should be unimportant. This paper
indeed states in its conclusions (I have not read the whole text), that both measure only disagree 2% of the time. information gain ratio vs gini index
Information Gain Ratio (formula see e.g. here: http://en.wikipedia.org/wiki/Information_gain_ratio
) adds another factor to penalize attributes with too many different values (in case of discrete attributes). Two many values mean that the base of every entropy calculation is rather small (on the average). Who do you trust more ? 1/2 or 5/10 ?
So the resulting interesting question is whether a gini index-ratio would perform as well as information gain ratio. I am pretty sure that a paper about that has already been written and forgotten