The traditional problem of influence maximizing aims to select a certain number of source seeds to publish specific information, so as to maximize the influence spread of the information. However, seed users selected by algorithms may not necessarily be willing to publish the specified information. In addition, traditional influence maximization algorithms need to be rerun on networks with different structures, resulting in lower efficiency. To address these issues, this paper first formalized the problem of maximizing influence as a new Key User Identification for Information Propagation (KUIP) problem, which involved how to discover a certain number of key users without requiring them to publish specific information, but by intervening in their attitudes and tendencies towards disseminating information, to maximize the influence spread of that information. In order to more accurately describe the information propagation scenario, this paper proposed an Adjustable Threshold Model (ATM) to simulate the attitude tendencies and environmental influences of users in disseminating information. Furthermore, in order to ensure the efficiency and effectiveness of key user identification on networks with different structures, this paper proposed a Key user identification based on Propagation Reinforcement Learning (KPRL) , which utilized graph attention mechanism to learn the propagation features of users and trained the model parameters using Double Deep Q-Network (DDQN) . Experiments on six real network datasets showed that KPRL improved the influence spread indicator by an average of 11.7%, surpassing existing baseline methods and demonstrating its effectiveness in the field of key user identification.