In this paper we devote to study some feature selection of an information system in which redundant or insignificant attributes in data sets can be eliminated. An approach of importance gain function is suggested to evaluate the global average information gain associated with a subset of features. A heuristic algorithm on iterative criterion of feature selection on the significance of attributes is proposed to get the least reduction of attribute set in knowledge discovery. The feasibility of feature selection proposed here is validated by some of examples.