<p>Mobile Edge Computing (MEC) has been proposed for its ability to provide computing service to users more efficiently, and a reasonable offloading strategy is especially critical to improve the performance of MEC. However, the dramatic increase in the number of MEC users leads to network congestion, and it is a challenge to make reasonable offloading decisions under missing data in MEC. On the other hand, the dynamics of the edge network and the varying preferences of applications increase the difficulty of making effective offloading decisions. In this paper, we propose a partial task offloading scheme based on state-action-reward prediction (PTOSAR) under missing data in MEC, which integrates user mobility, task cost diversity, and task information missing probability. Given the time-varying nature of the network environment, we formulate the task offloading problem as a Markov Decision Process (MDP) whose goal is to minimize the soft-user cost (SUC). To solve the missing data problem in MEC, a sequence prediction model integrating state-action-reward information using the transformer architecture is established to further strengthen the relevance of state representation and action representation to the deep learning reinforcement (DRL) strategies. PTOSAR incorporates an action prediction model and a reward prediction model as additional learning constraints to guide the learning of state and action representations. Simulation results show that the proposed algorithm converges well and outperforms existing benchmark algorithms regarding decision rationality and SUC.</p>