Automated facial expression recognition can greatly improve the human-machine interface. The machine can provide better and more personalized services when it knows the human's emotion. This kind of improvement is an important progress in this artificial intelligence era. Many deep learning approaches have been applied in recent years due to their outstanding recognition accuracy after training with large amounts of data. The performance is limited, however, by the specific environmental conditions and variations in different persons involved. Hence, this paper addresses the issue of how to customize the generic model without label information from the testing samples. Weighted Center Regression Adaptive Feature Mapping (W-CR-AFM) is mainly proposed to transform the feature distribution of testing samples into that of trained samples. By means of minimizing the error between each feature of testing sample and the center of the most relevant category, W-CR-AFM can bring the features of testing samples around the decision boundary to the centers of expression categories; therefore, their predicted labels can be corrected. When the model which is tuned by W-CR-AFM is tested on extended Cohn-Kanade (CK+), Radboud Faces database, and Amsterdam dynamic facial expression set, our approach can improve the recognition accuracy by about 3.01%, 0.49%, and 5.33%, respectively. Compared to the competing deep learning architectures with the same training data, our approach shows the better performance.