The problem of object generalization with account for the necessity of processing the incomplete and inconsistent information stored in real databases is considered. It is suggested to use means of rough sets theory and decision trees to generalize the information stored in real databases. Noise models are presented, and a noise effect on the operation of generalization algorithms using the methods of building decision trees is developed. The algorithm for unknown values reconstruction in learning samples subjected to the noise effect based on the nearest neighbour method is proposed. The results of program modeling are brought out.