For the refinement of the database in data mining,by analyzing the characteristics and shortcomings of the current attribute reduction methods,and combining with it the features of simple operation and rapid classification of decision tree,the authors established a kind of attribute reduction method (BD-RED) based on decision tree using rule description of the knowledge and similarity measures between rules families.Further,we discussed the explanatory construction of similarity measure between rules families,gave the specific implementation strategy of BD-RED,and analyzed the performance through examples.The results show that BD-RED has a good structure and strong operability,and is an effective way to achieve attribute reduction under different consciousness,so it can be suitable to the large-scale attribute reduction.