As an emerging research topic for proximity service (ProSe), automatic emotion recognition enables the machines to understand the emotional changes of human beings which can not only facilitate natural, effective, seamless, and advanced human–robot interaction or human–computer interface but also promote emotional health. Facial expression recognition (FER) is a vital task for emotion recognition. However, significant gap between human and machine exists in FER task. In this paper, we present a conditional generative adversarial network-based approach to alleviate the intra-class variations by individually controlling the facial expressions and learning the generative and discriminative representations simultaneously. The proposed framework consists of a generator G and three discriminators ( Di , Da , and Dexp ). The generator G transforms any query face image into another prototypic facial expression image with other factors preserved. Armed with action units condition, the generator G pays more attention to information relevant to facial expression. Three loss functions ( , La , and Lexp ) corresponding to the three discriminators ( Di , Da , and Dexp ) were designed to learn generative and discriminative representations. Moreover, after rendering the generated expression back to its original facial expression, cycle consistency loss is also applied to guarantee the identity and produce more constrained visual representations. Optimized by combining both synthesis and classification loss functions, the learnt representation is explicitly disentangled from other variations such as identity, head pose, and illumination. Qualitative and quantitative experimental results demonstrate the proposed FER system is effective for expression recognition.