High-dimensional data modeling in the fields of machine learning and pattern recognition is ubiquitous and of great value. The ``curse of dimensionality" problem that exists in the high-dimensional data analysis process constrains the effective intervention of many machine learning models. Many effective methods for subspace and non-negative matrix reconstruction have been proposed. Non-negative matrix reconstruction can improve algorithm construction in unsupervised and semi-supervised learning by improving the loss function and adding priors. This paper proposes a non-negative matrix factorization loss function based on adaptive dual-weight learning. The proposed loss-function learns based on the class structure information of the data set in high-dimensional space and low-dimensional space, uses weighted $L_{2,1}$ norm to improve model robustness, and uses weighted learning strategies to learn approximation in low-dimensional space. This results in better algorithmic robustness. Experimental results on some benchmark datasets and hyperspectral images demonstrate the superiority of the new algorithm.