This study benchmarks whether GeoShapley, a game-theoretic explainer, can recover spatially varying coefficients from machine learning models using location encoder embeddings. Eleven encoders from the TorchSpatial framework were evaluated against a synthetic process with known coefficients across grid, county, and global scales.
- The benchmark tested eleven encoders under untrained and contrastively trained conditions, both with and without raw coordinates.
- Recovery of the primary coefficient was consistently high across all encoders.
- Recovery of the secondary coefficient was scale-dependent, showing the most variation at the global scale.
- The raw-coordinate baseline remained competitive throughout the evaluations.
The results indicate that while location encoder embeddings can effectively capture primary spatial effects, their ability to recover secondary coefficients varies significantly by scale.