Intuitive Overview of the Algorithm
Most common feature selection algorithms are primarily oriented toward favoring features that are at least somewhat predictive over the entire domain of the feature set. This predictivity may be nonlinear, and it may interact with other features, but such a predictor will be at a significant advantage over more powerful but only locally predictive candidates if the nature of its relationship to a target variable is at least somewhat consistent across the domain of all possible values of all candidate features.
This global favoritism ...