Based on high noise and redundancy of gene expression data and that traditional NMF algorithm is inefficient in the clustering of gene expression data， a new clustering method of beta divergence matrix decomposition under the constraint of smooth lo norm and the combination Kmeans is presented， and the new clustering method is applied to gene expression data. The smooth lo norm is introduced into the objective function of matrix decomposition based on beta divergence so as to extract the useful feature information for the clustering. Finally， compared by experiments， the average clustering accuracy of the improved algorithm reaches 70 percent， which is 11 percent higher than that of the traditional NMF clustering algorithm， and clustering effect is more significant than other methods.