Nonnegative matrix factorization(NMF) is an effective dimension reduction method, which is widely used in image clustering and other fields. Some NMF variants preserve the manifold structure of the original data. However, the construction of the traditional neighbor graph depends on the original data, so it may be affected by noise and outliers. Moreover, these methods are unsupervised and do not use available label information. Therefore, this paper presents an adaptive graph-based discriminative nonnegative matrix factorization(AGDNMF). AGDNMF uses the available label to construct the label matrix, such that the new representations with the same label data are aligned to the same axis. And the neighbor graph in AGDNMF is obtained by adaptive iterations. A number of experiments on many image data sets verify that AGDNMF is effective compared with the other state-of-the-art methods.

非负矩阵分解（NMF）是一种有效的降维方法，广泛应用于图像聚类等领域。一些NMF变体保留了原始数据的多种结构。但是，传统邻居图的构造取决于原始数据，因此可能会受到噪声和离群值的影响。此外，这些方法不受监督，并且不使用可用的标签信息。因此，本文提出了一种基于自适应图的判别式非负矩阵分解法（AGDNMF）。AGDNMF使用可用的标签构造标签矩阵，以使具有相同标签数据的新表示与同一轴对齐。通过自适应迭代得到AGDNMF中的邻居图。