基于Sinkhorn距离特征缩放的多约束非负矩阵分解算法

为了减少原始特征对非负矩阵分解(NMF)算法的共适应性干扰，并提高NMF的子空间学习能力与聚类性能，该文提出一种基于Sinkhorn距离特征缩放的多约束半监督非负矩阵分解算法。首先该算法通过Sinkhorn距离对原始输入矩阵进行特征缩放，提高空间内同类数据特征之间的关联性，然后结合样本标签信息的双图流形结构与范数稀疏约束作为双正则项，使分解后的基矩阵具有稀疏特性和较强的空间表达能力，最后，通过KKT条件对所提算法目标函数的进行优化推导，得到有效的乘法更新规则。通过在多个图像数据集以及平移噪声数据上的聚类实验结果对比分析，该文所提算法具有较强的子空间学习能力，且对平移噪声有更强的鲁棒性。

Multi-constrained Non-negative Matrix Factorization Algorithm Based on Sinkhorn Distance Feature Scaling

In order to reduce the co-adaptability interference of the original feature to the Non-negative Matrix Factorization (NMF) algorithm and improve the performance of non-negative matrix factorization subspace learning and clustering performance, a novel multi-constrained semi-supervised non-negative matrix factorization algorithm based on Sinkhorn distance feature scaling is proposed. First, the algorithm is feature-scaled by the Sinkhorn distance to the original input matrix to improve the correlation between features of the same type of data in the space, then, the dual graph manifold structure combined with the sample label information and the norm sparsity constraint are embedded in the model as a dual regular term, so that the decomposed base matrix has sparse characteristics and strong spatial expression ability. Finally, the objective function of the proposed algorithm is optimized by Karush-Kuhn-Tucker (KKT) conditions, and effective multiplication update rules are obtained. Through the comparative analysis of the results of multiple clustering experiments on multiple image data sets and translational noise data, the algorithm proposed in this paper has a strong subspace learning ability and is more robust to translational noise.