In machine learning, datasets are often disturbed by different noises. The Nonnegative Matrix Factorization (NMF) algorithm provides a robust method to deal with noise, which will significantly improve the efficiency of machine learning. In this investigation, the standard NMF algorithm and L2,1-Norm Based NMF algorithm are studied by designing experiments on different noise types, noise levels, and datasets. Furthermore, Relative Reconstruction Errors (RRE), accuracy, and Normalized Mutual Information (NMI) are used to evaluate the robustness of the two algorithms. In this experiment, there is no significant difference in performance between the two algorithms, while L2,1-Norm Based NMF algorithm shows relatively small advantages.