The formulation of the prediction model to compensate for the thermal error of machine tools is a common method to effectively solve the decline of machine tool accuracy caused by thermal error. This article proposes an adaptive robust modeling method for thermal error of computer numerical control machine tools based on regularization, which can adaptively select temperature-sensitive points (TSPs) in the modeling process, and has high prediction accuracy and robustness. Firstly, the robustness mechanism of thermal error modeling is analyzed, which is based on the principle of structural risk minimization. Secondly, the sparsity of the solution of least absolute shrinkage and selection operator (LASSO) in regularization algorithms is used to realize adaptive TSP selection. Then, based on the thermal error data under different experimental conditions, the prediction effect of the proposed modeling method is analyzed and compared with the commonly used multiple linear regression, back propagation (BP) neural network, and ridge regression algorithms. Results show that the proposed modeling method has the highest prediction accuracy and robustness, which are 5. 22 and 1. 69 μm, respectively. Finally, the thermal error compensation experiment is implemented by using the established prediction model to evaluate the actual compensation effect of the proposed modeling method.