Abstract: In machine learning and artificial intelligence，rough set theory is an important theory for dimension reduction.However，one given information system may have more than one reduct，which will lead to different knowledge.Therefore，it is a critical problem to select one suitable reduct for knowledge discovery.Aiming at this，this paper appends additional information——preference relation into information system and uses it to guide the procedure of finding an reduct，which will produce optimal reduct under preference.Also，this paper proves the correctness of the algorithm and uses it to reduce the information system.In subsequent experiment，this paper applies the optimal reduct in classification problem and receives good effect.