Contemporary robot navigation systems are not only expected to ensure stability but also to prioritize comfort and social interaction, such as maintaining an appropriate distance from pedestrians and adeptly maneuvering through crowds. Traditional frameworks often fall short in addressing navigation challenges within the context of human-robot interaction, merely treating nearby individuals and objects as obstacles. To tackle this issue, we propose a graph-based architecture for active reinforcement learning. This approach visualizes human-robot, human-human, and robot-robot interactions through a graph representation. An active policy, learned through active learning, adjusts the agents’ perspectives and serves as an explorer to enhance the performance of our proposed method. Ultimately, this graph-based active reinforcement learning architecture is refined through reinforcement learning. Various crowd environments have been established to validate the effectiveness of our proposed method. Simulation experiments demonstrate that compared to traditional graph neural networks, reinforcement learning methods, and active learning methods, the Graph Neural Network + Active Learning approach is superior.