<p>The rapid evolution of wireless communication necessitates the development of advanced channel modeling techniques to address the increasing complexity of next-generation networks, including 6G and beyond. Conventional techniques primarily employ empirical and stochastic approaches, which pose limitations in adapting to varying environments and propagation scenarios. A new hybrid graph-attentive neural framework (HGANF) for intelligent wireless channel modeling is proposed, in which the apex intertwined structure for spatial–temporal representation learning is achieved through the combination of the graph neural networks (GNNs) and self-attention mechanisms. While conventional ML-based channel models mainly perform feature extraction from raw measurement data, the proposed framework refers to the iterative construction of a feature vector that dynamically represents the non-Euclidean topological relation in a multipath propagation environment. Thus, the HGANF model adopts the graph embeddings to express these nonlinear interactions of complex signals and an adaptive attention mechanism to selectively prioritize the channel parameters in real time. Both simulation and field-tested experiments illustrate that HGANF is superior to the state-of-the-art deep learning methods with respect to accuracy, generalization, and computational efficiency, achieving 32% less estimation error than current deep learning-based methods. The proposed framework is also more robust in various propagation environments such as urban microcells, high-mobility vehicular networks, and mmWave environments. These contributions distinguish HGANF from prior GNN-attention hybrids by incorporating a real-time graph updating mechanism responsive to signal dynamics (e.g., delay, Doppler, power), a custom adversarial robustness layer to counter perturbation-based attacks, and a lightweight design suitable for deployment in 6G use cases.</p>