MamBRA: session-level bandwidth prediction for adaptive video streaming using selective state space models

Live streaming is the real-time transmission of video content to an audience as it is simultaneously recorded. This technology is frequently utilized for applications such as covering live events and facilitating video calls. By dynamically modifying the video quality to match network conditions and device capabilities, adaptive video streaming provides improved Quality of Experience (QoE). However, as user demands for high quality and low latency increase, using efficient video streaming systems is getting harder. In addition to taxing network resources, the increase in video traffic is lowering video quality. Deep and transformer learning algorithms use data-driven methods to optimize video delivery, enhance QoE, and lessen network congestion in order to overcome these obstacles. Mamba utilizes the efficient linear complexity of selective state space model (SSM) mechanism to process data sequences more effectively. This paper proposes an adaptive video streaming framework (MamBRA) based on Mamba for session-level bandwidth prediction. The model is trained in a supervised time-series manner on disjoint user sessions to prevent information leakage and preserve temporal structure. During inference, it leverages the linear state-space formulation of Mamba to efficiently generate stable bandwidth predictions within each session. Experimental results demonstrate reduced prediction error, improved accuracy, and enhanced temporal stability. The model achieves an overall inference accuracy of 93.94%, with session-level accuracy reaching as high as 97.32%. Furthermore, the predicted bandwidth achieves more consistent QoE scores compared to the PPO-based approach used in Pensieve.