A multi-agent proximal policy optimization method based on a causal hypergraph neural network (MAPPO-HGNN) was proposed to address the performance degradation caused by overlapping link interference and rapidly time-varying channels in high-speed vehicular networks, with the aim of improving network throughput, latency performance, and resource utilization. A causal hypergraph was constructed to characterize the joint interference structure arising from resource sharing among multiple links. A hypergraph neural network based on spectral-normalized propagation was designed to obtain structured representations of high-order interference relationships. The resulting hypergraph embedding features were subsequently integrated into the multi-agent reinforcement learning framework, enabling adaptive spectrum resource allocation. Simulation results showed that, compared with three baseline algorithms, the proposed method achieved faster convergence and greater stability under different vehicle densities. Specifically, the average throughput and resource reuse ratio were increased by 25.2% and 15.1%, respectively, while the task latency was reduced by 16.7%. These results validated the effectiveness and robustness of the proposed method in high-speed, non-stationary vehicular networking scenarios.