In this paper, a novel and effective algorithm is introdcued for constructing decision tree. First of all, the knowledge dependence in rough set theory is used to reduce the test attribute set of decision tree, that is, the test attribute space is optimized and hence the attributes which are not correlated with the decision information are deleted. Then in view of the shortcomings existing in ID3 algorithm, the degree of dependency of decision attribute on condition attribute is used as a heuristic information for selecting the attribute that will best sepatate the samples into individual classes. Thus the repetition of the decision subtrees and some attributes to be chosen many times on the same decision tree are resolved. The example shows that the method is better than the ID3 algorithm and has been verified to be effective.