Why filtering by gain ratio rather than correlation? I'm new to the filtering problem in data mining. Suppose I've 100000 numeric features I can use to predict 1 nominal variable. Suppose I want to use only the best 5 features. Why shall I not use the 5 features with the highest Spearman rho? I guess the answer is the possible existence of co-correlation among the 5 features with the highest rho. Is this what info gain or gain ratio solves? 
 A: Yes there might be correlation among the five best predictors which may not properly predict the outcome and results in insignificance of a predictor  but there is another catch also due to which selecting the predictors with high spearman value w.r.t outcome is not the right way.
This catch is the interaction effect, which tells us that even if we have two predictors which on their own are not providing a good correlation with the outcome, but in combination they are fitting with the outcome very well. To give you an example suppose we are observing the effect of advertising on sale of our product and we are observing whether ads on TV, Social Media, WebPages, radio or newspaper is effecting our sales and we want to spend the sales budget so that it is most efficient and contribute to the highest sales. There may the case that TV and radio alone are not very good for the sales but combining TV and radio sales gives us the biggest boost. In this case if we had only gone with the spearman value we would have not selected TV and radio due to their low correlation value wrt to sales but combining these two(which is called interaction effect) is the best we can do to maximise our sale.
