Most community detection methods are aiming at solving undirected and unweighted datasets. However, datasets are often directed and weighted with noisein real world. In order to process noisy and directed weighted community detection, a method based on nonnegative matrix factorization (NMF) is proposed. In the algorithm, wavelet threshold denoising is used to denoise the social network datasets. And the community structure is abtained by community detection through NMF. Simulations show the proposed method is more effective,i.e. for Lesmis dataset when SNR is 15, the accuracy of dividing community is 96% and the modularity of the method is improved by 29%. The proposed method is more applicable than other community detection methods for directed weighted datasets with noise.