Feature selection is one of important research areas of data mining. In real world, many systems to be investigated are very complicated, some of them have problems of noise and information-absent. Through several feature selection methods——evaluation of samples classification and evaluation of each element of feature set, some features with great influence to the performance of the complex information-absent system can be found. Then a practical and correct direction to improve this system is pointed out. The improvement has been proved by the experimental results.