Propagation Graph Representation Learning and Its Implementation in Direct Path Representation

This paper proposes a novel graph-learning-based radio propagation model, referred to as propagation graph representation learning. Recent advancements in deep learning have succeeded in developing site-specific path loss prediction; however, predicting shadowing without on-site measurement data is still a critical challenge. Propagation graph representation learning aims to express the site-specific propagation process. In the envisioned graph, nodes represent a transmitter, receivers, and obstructions, while edges represent propagation conditions between nodes. The graph structure enables us to recognize the reflection, diffraction, and shielding for accurate shadowing estimation. We also propose a feedforward neural network (FFNN) based representation model in a direct path scenario. Through the simulation using actual datasets in urban areas, we demonstrate that the proposal achieves twenty times faster computation than ray tracing while predicting shadowing well.