Abstract
<title>Abstract</title> <p>Feature importance from explainable machine learning increasingly guides regional analysis and policy advice. Existing cautions establish what importance is not: not a causal effect, not a policy lever. But they were developed in aspatial settings, and they leave a prior question open. Is an importance ranking stable across the spatial regimes that regional processes exhibit? We test that question on internal migration across 229 South Korean municipalities (2017–2024). We partition the attributions of a single fitted gradient-boosting model by urban development stage and evaluate the resulting rank differences against a permutation null obtained by reassigning stage labels across municipalities. Twelve of twenty predictors are stage dependent. Stage dependency, however, is largely unrelated to global importance. The four leading predictors move by at most four rank positions, whereas the largest re-orderings occur in the middle of the ranking. Magnitude also fails to disclose direction. Network prominence is the most informative predictor in sparsely populated municipalities, yet its contribution there is negative: the model relies on the absence of network position. Childcare provision, the leading predictor overall, reverses the sign of its contribution across the settlement hierarchy. Cross-model comparison, cluster bootstrapping, and spatial transfer tests indicate that these patterns are not artefacts of model choice. Feature importance in spatially heterogeneous systems is therefore best read as evidence about the spatial organization of predictive information. Its direction and its spatial uniformity must be established separately, because the ranking encodes neither.</p>