The traditional attribute reduction algorithms such as rough set will fail to get accurate results when deal with
the data sets which have noise or labeling errors. Therefore, this paper proposes an attribute reduction algorithm which can
analyze this kind of data effectively. Firstly, the best projection of the data sets is obtained by using the maximum margin
projection(MMP) method. Then ??2,1-norm on the projection matrix is used to achieve row-sparsity, which leads to selecting
relevant features. Finally, the proof of the algorithm’s convergence and validity to the data sets with errors is given. The
result of experiments on the UCI data sets show the effectiveness of the proposed algorithm.