The surge in wireless connectivity demand, coupled with the finite nature of spectrum resources, compels the development of efficient spectrum management approaches. Spectrum sharing presents a promising avenue, although it demands precise characterization of spectrum demand for informed policy-making. This paper introduces HR-GAT, a hierarchical resolution graph attention network model, designed to predict spectrum demand using geospatial data. HR-GAT adeptly handles complex spatial demand patterns and resolves issues of spatial autocorrelation that usually challenge standard machine learning models, often resulting in poor generalization. Tested across five major Canadian cities, HR-GAT improves predictive accuracy of spectrum demand by 21% over eight baseline models, underscoring its superior performance and reliability.
Figures & tables
Fig. 1: Illustration of the HR-GAT hierarchical graph construction.
Feature Category
Description
Population Density
Number of residents per grid tile
Household Density
Number of households per grid tile
Daytime Population
Estimated population during working hours
Economic Activity
Number of businesses by sector
Infrastructure Density
Number of buildings and their total area
Road Network
Total road length and number of roads
TABLE I: Summary of Features Processed for Demand Modeling
Fig. 3: Deployed bandwidth heatmaps at zoom level 15 for Calgary, Montreal, Toronto, Vancouver, and Ottawa. Darker regions indicate higher estimated spectrum deployment.
Proxy Variable
R²
F-Statistic
p-Value
Deployed Bandwidth (PBW)
0.727
4477
<0.001
TABLE II: OLS Regression Results for Proxy Validation
Model
Median MAE ↓
Median RMSE ↓
R² ↑
Gradient Boosting
23.44
44.27
0.86
XGBoost
23.53
44.72
0.86
LightGBM
30.78
55.70
0.80
Decision Tree
41.60
79.92
0.53
Random Forest
45.89
77.48
0.55
Linear Model
55.50
89.31
0.43
TABLE III: Performance Comparison using CBCV
Fig. 4: eCDF of RMSE values across models. HR-GAT demonstrates superior prediction accuracy.
Fig. 5: Log-scaled residual error distribution across cities, showing HR-GAT’s consistent performance.
Model
Median MAE (Ottawa) ↓
Gradient Boosting
37.34
XGBoost
39.49
LightGBM
47.20
Decision Tree
52.47
Random Forest
50.66
Linear Model
60.93
TABLE IV: Median MAE for Ottawa (Unseen Test Set)
Fig. 6: SHAP summary plot illustrating the relative importance of the top 10 features in predicting mobile spectrum demand. Higher absolute SHAP values indicate stronger influence on the model’s estimation.