<p>Compared to traditional video streaming, 360° video streaming consumes significantly more bandwidth. To address this, tile-based methods have been proposed to reduce bandwidth consumption and enhance the quality of experience (QoE) for users. These methods employ viewport prediction algorithm to select tiles containing the user's region of interest and utilize bitrate allocation algorithm to determine appropriate bitrates for different tiles. However, current viewport prediction algorithms insufficiently utilize viewport spatial position information, affecting prediction accuracy. Additional tile transmissions caused by prediction errors result in bandwidth wastage. Furthermore, the spatial offset of the viewport directly influences tile transmission throughput, making accurate viewport prediction essential for enhancing the performance of bitrate allocation algorithms. Existing bitrate allocation algorithms primarily focus on maximizing QoE for a single user. In the practical environment, factors such as varying viewports among users and time-varying network features can lead to unbalanced QoE distribution among users, making it challenging to ensure QoE fairness among multiple users. To tackle these issues, we propose an algorithm combining Spatio-Temporal GCN-based viewport prediction and Multi-agent deep Reinforcement learning-based bitrate allocation for users’ QoE-fairness (STMRQ). Firstly, a viewport prediction algorithm based on spatio-temporal graph convolutional network is introduced, which extracts the spatio-temporal features of video tiles through a novel viewport position matrix, thereby improving the accuracy and computational efficiency of viewport prediction. Subsequently, a 360° video streaming bitrate allocation algorithm based on multi-agent reinforcement learning is proposed, modeling the environment of multi-user dynamic bandwidth streaming on demand. By leveraging state parameters such as user viewport variability, client status, and global network features, we employ a multi-agent reinforcement learning algorithm to train bitrate allocation strategies in a multi-user environment. Extensive experimental evaluations demonstrate that, compared to existing viewport prediction methods and bitrate allocation algorithms, the proposed STMRQ achieves a higher average QoE across common datasets and effectively ensures QoE fairness. Additionally, the accuracy of viewport prediction is improved by up to double, with relatively low computational overhead during training.</p>