The measure of attribute selection is a heuristic in nature for selecting the splitting criterion that “best” separates a given data partition, D, of class-labeled training tuples into individual classes.
It determines how the tuples at a given node are to be split.