Regression is a fundamental problem in machine learning. While existing regression methods mainly suffer from low prediction accuracy or poor interpretability, in this paper, we propose a new regression framework named Decomposed Squeeze-and-Excitation Transformer (DSEFormer), which consists of Matrix Decomposition Attention Mechanism (MDAM) and the Feature Calibration Module (FCM). MDAM calculates the attention coefficients through multiple attention modules and utilizes PLU matrix decomposition to calculate the attention weights to obtain weighted feature representations. The FCM module consists of multilayer perceptron (MLP), convolution operation, and squeeze-and-excitation modules to calibrate the weighted features to improve the learning performance of the model. Feature weights improve the interpretability of the predicted results. Integration of MDAM and FCM promotes prediction accuracy. We perform an extensive experimental evaluation using public datasets and our own constructed movie and painting datasets. Ablation studies show that MDAM and FCM improve the prediction performance, e.g., MSE increases by 0.9802 and 1.0957 after removing FCM and MDAM, respectively. Moreover, we compare DSEFormer with three machine learning algorithms and six deep learning algorithms; the experimental results show that DSEFormer’s prediction error is lower than the competitors.