Abstract: Multiple-target regression refers to a set of input variables to estimate its corresponding multiple continuous attribute values, which has a wide range of applications in the field of data mining. The current research on multi-target regression tasks is based on the assumption that the label value is accurate. However, in actual situations, some labels of the data set may not be accurate, that is, some labels have certain noise. In this case, the traditional multi-target regression methods usually cannot achieve good results. In order to solve the multi-target regression problem in the above situation, the characteristics of large data samples in big data are used to refine the correlation between labels, and then the correlation matrix between labels is used to reconstruct the labels. Since the number of labels in the multi-target problem is usually large, the noise interference of some labels can be diluted to a certain extent. In addition, low-rank matrix decomposition is used to establish a mathematical optimization problem for the above ideas, and on this basis, kernel techniques are introduced to improve the nonlinear fitting ability of the model. Finally, a non-convex approximation method is used to solve the optimization problem, thereby ensuring the prediction performance of the multi-target regression model. The experiment is compared with 6 existing multi-target regression models on 18 datasets. The method proposed in this paper has obvious performance advantages in scenarios with large sample sizes.