Research on Graph Learning-Based Co-Inference Method for Space-Air-Ground Integrated Networks

Abstract

In the space-air-ground integrated network (SAGIN), diverse artificial intelligence (AI) applications are realized through the inference of pre-trained deep neural networks (DNNs). Owing to the ever-increasing parameter scale of DNNs, the execution of inference tasks under constrained communication and computational resources typically necessitates multi-party collaboration. This paper integrates cognitive radio (CR) technology into SAGIN, aiming to minimize inference latency while considering constraints such as satellite communication interference thresholds, device quality-of-service requirements, and transmit power limitations. We formulate a cooperative inference problem involving device association, model splitting, and power optimization, with unmanned aerial vehicle (UAV) serving as task initiators. The considered problem is a non-convex optimization with coupled multi-dimensional variables. Accordingly, we decompose the original problem based on optimization variables and propose a graph learning method combining quadratic unconstrained binary optimization (QUBO) and an attention mechanism. Specifically, the objective function and constraints are modeled in QUBO form, and a differentiable loss function is derived through Hamiltonian relaxation, where a dynamic multi-head attention mechanism is then employed to aggregate node information in the graph-structured problem, optimizing device association and model splitting. Additionally, transmit power is optimized using successive convex approximation (SCA). Ultimately, the three sub-problems are solved alternately in an iterative process. Simulation results indicate that the proposed method reduces the total latency of collaborative inference tasks by 11%.