Throughput Estimation of OBSS WLANs Using Ensemble Graph Neural Networks

Predicting throughput in dense and dynamic Wireless Local Area Networks (WLAN) deployments is a critical challenge hampered by complex protocols and interference. Traditional methods struggle with high runtime or significant prediction errors, hindering their practical applications. This paper proposes an approach employing Graph Neural Networks (GNNs) for accurate throughput estimation in Overlapping Basic Service Sets (OBSS) within WLANs. Leveraging signal strength, airtime, Signal-to-Interference + Noise Ratio (SINR), and interference levels as features, the proposed method utilizes GNN models like GraphSAGE, Graph Convolution Networks (GCN), (Graph Attention Networks (GAT), and ChebNet. An ensemble stacking model with a random forest meta-model is introduced to enhance prediction accuracy. Addressing the decentralized nature of IEEE 802.11 networks, our approach enables reliable throughput projections for optimized network design. This study paves the way for improved WLAN efficiency and reliability by facilitating real-time throughput predictions with minimal error. Our dataset and proposed stacking model achieve an impressive R2 score of 0.957, demonstrating its effectiveness and promising potential for real-world applications.