Attributes play a significant role in data mining and machine learning. In real-world problems, it is reasonable to consider different attributes with different importance. Most of the existing attribute importance research focuses on statistics or information theory perspectives, that is, an attribute with a higher prediction ability is considered to be more critical. User preference, which involves a user to specify his or her preferential attitude towards a set of attributes, is also a meaningful perspective in attribute importance analysis. However, this field has not received its due attention. Three-way decision theory provides a unified framework for thinking, problem solving, and information processing in three. For the purpose of analyzing attribute importance based on user preference, three-way decision is introduced as a framework in this thesis. We represent user preference of attributes in qualitative and quantita- tive ways. For the qualitative analysis, by using TAO (Trisecting-Acting-Outcome) model in three-way decision as a framework and binary relations as concrete method, we rank attributes by considering their importance. For the quantitative analysis, we use three-level computing model in three-way decision as a framework and eigenvector method for concrete computing, we calculate numerical weights for each attribute as its importance. Finally, we use evaluation based three-way analysis to categorize the results of qualitative and quantitative analysis into different groups of importance.