Current neighborhood rough sets have been usually used to solve complete information system, not incomplete system. In order to solve this problem, an extended tolerance relation was proposed to deal with the incomplete mixed information system, and associative definitions were provided. The degree of complete tolerance and neighborhood threshold were used as the constraint conditions to find the extended tolerance neighborhood. The attribute importance of the system was got by the decision positive region within the neiborhood, and the attribute reduction algorithm based on the extended tolerance relation was proposed, which was given by the importance as the heuristic factor. Seven different types of data sets on UCI database was used for simulation, and the proposed method was compared with Extension Neighborhood relation (EN), Tolerance Neighborhood Entropy (TRE) and Neighborhood Rough set (NR) respectively. The experimental results show that, the proposed algorithm can ensure accuracy of classification, select less attributes by reduction. Finally, the influence of neighborhood threshold in extended tolerance relation on classification accuracy was discussed.