Abstract: As a classical manifold learning algorithm, local linear embedding(LLE)can obtain low-dimensional manifolds in a high-dimensional space. However, it is sensitive to the selection of neighbor samples and lacks the ability of global structure retention. To solve this problem, we propose an improved LLE algorithm. Considering the differences among samples and the global representation of data, we improve the information mining performance of the algorithm in the process of dimensionality reduction by introducing a discreteness retention term and global weighting index; moreover, we reduce the sensitivity of the algorithm to noise and overcome the defect that the traditional LLE algorithm only focuses on the local manifold features and ignores the global structure. A numerical simulation and a soft sensor application of wheat grain protein content verify the effectiveness and superiority of the algorithm.