Answers are grounded in this paper and cite their sources.
Beacon towers were warning and communication nodes in the Ming Great Wall defense system. We developed an interpretable graph learning framework combining a graph attention network (GAT) with random forest (RF) for Yansui Garrison. The GAT learned graph embeddings by integrating 11 node variables with spatial neighborhood relations, and RF used these embeddings to predict relative siting propensity. Connection weights, component correlations, and partial dependence plots characterized how siting relationships differed between individual and neighborhood scales. GAT-RF achieved a test R ² of 0.896, outperforming RF on original variables (0.842) and GAT alone (0.561), with stable performance across distance thresholds. Individual modes declined from east to west, while neighborhood modes increased. This shift reflected route differences between individual and neighborhood scales. Beacon tower siting can be understood in terms of a scale-dependent spatial logic, in which node attributes and neighborhood relations are jointly associated with the spatial organization of networked heritage.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: