Decision tree algorithm has been widely used to classify numeric and categorical attributes. Lots of approaches were suggested in order to induce decision trees. ID3 (Quinlan, 1986), as a heuristic algorithm, is very classic and popular in the induction of decision trees. The key of ID3 is to choose information gain as the standard for testing attributes. In this paper, we propose a novel measure based on rough set theory to select attributes that will best split current samples into individual classes. In the view of rough set theory, we analyze the shortcomings of ID3 algorithm and rationality of the new approach, and then propose a fixed algorithm based on original idea. The results of example and experiments show that our approach is better in selecting nodes for inducing decision trees than ID3.