<![CDATA[Data-driven studies of coffee shop location are usually validated through random splits and interpreted through global importance rankings, leaving their accuracy estimates and explanations open to question. This study aims to compare tree-based ensembles with a linear baseline under spatially blocked validation, to quantify feature contributions, and to interpret the co-location structure of coffee shops in Surabaya, Indonesia. Across a hexagonal grid, 436 cells split evenly between presence and pseudo-absence were described by eleven OpenStreetMap-derived spatial features. Random Forest, XGBoost, and a logistic regression baseline were evaluated under nested cross-validation blocked by the 31 sub-districts, since the labels were significantly autocorrelated, with paired significance testing and bootstrap confidence intervals. The ensembles proved statistically equivalent, and neither surpassed the linear baseline in ranking ability, with areas under the receiver operating characteristic curve of 0.865, 0.862, and 0.867. They improved threshold-dependent metrics instead, raising the Matthews correlation coefficient from 0.493 to 0.562 and recall from 0.715 to 0.814. Hotel and competitor density led the permutation importance ranking, with no ablated configuration below 0.842. Shapley additive explanations exposed a saturating co-location structure in which competitor density contributed positively to presence, a pattern linear models cannot represent. Spatial blocking altered the metrics by less than 0.01, guarding against over-optimistic evaluation rather than driving these results, and the feature profiles give operators and municipal planners an instrument for screening candidate areas.]]>
Copyrights © 2026
