Cyber-physical-social systems (CPSS) combining cyber-physical systems (CPS) with social networking aim to provide the personalized big data services for humans. To support CPSS, cloud is usually explored to store the sensing data from mobile ends. However, the high quality-of-service (QoS) of CPSS challenges the unstable and non-real-time access links between cloud and ends. To this end, distributed edge node, pushing the storage resources onto the network edge, caches the popular meaningful data, or contents from cloud. We are thus motivated to integrate cloud, edge and CPSS to construct cloud-edge assisted CPSS, i.e., CE-CPSS. To further improve the QoS, we consider the issue of multi-objective joint optimization, i.e., maximizing edge hit ratio while minimizing content access latency and traffic cost. To solve this complex problem, we focus on the deep reinforcement learning (DRL)-based method for proactive caching. Specifically, the distributional deep Q network (DDQN) seeks to build a distribution model on returns to optimize the policy of content decision-making. Fortunately, generative adversarial network (GAN) still concentrates on learning the data distribution and generating compelling data. In addition, the prioritized experience replay (PER) is helpful to learn from the most effective sample. So we construct a novel algorithm based on DDQN, GAN, and PER called PG-DDQN. The experiments prove that our proposal successfully outperforms the existing approaches.