Strategic Location and Survival Dynamics of Businesses in San Diego County: A Comparative Machine Learning and Geospatial Approach

Background: Municipal business registries provide valuable information for examining business conditions and characteristics associated with longevity. However, recorded business-status fields do not necessarily contain complete event histories required for formal survival analysis. This study therefore examines recorded active and closed/canceled business status rather than time-to-event survival probabilities.Purpose: This study compares logistic regression with four tree-based ensemble classifiers to distinguish recorded business status among San Diego County establishments. SHAP interpretation and unsupervised clustering are also used to examine structural and geographic patterns.Methods: The integrated dataset contained 158,330 initial records and 140,737 establishments after cleaning. The supervised modeling dataset included 137,814 observations after excluding 2,923 records without valid geographic coordinates. A stratified 80/20 split produced 110,251 training and 27,563 holdout test observations. SMOTE was applied exclusively to the training data, balancing both classes at 69,998 observations each. Five classifiers were evaluated on the same unaltered holdout set. SHAP was used to interpret LightGBM predictions, while K-Prototypes (k = 4) and spatial K-Means (k = 7) characterized business profiles and geographic patterns.Results: Logistic regression achieved an F1-score of 0.587087, accuracy of 0.661834, and ROC-AUC of 0.683970. LightGBM achieved the highest accuracy (0.835679), recall (0.760220), F1-score (0.792553), and ROC-AUC (0.911295). XGBoost achieved the highest precision (0.829190). SHAP identified business age as the strongest contributor to LightGBM predictions, followed by ownership, density, geographic, and industry features. Clustering identified distinct business profiles and spatial concentrations.Conclusions: Ensemble classifiers outperformed the logistic-regression baseline on the holdout dataset. Business age and ownership structure contributed substantially to model predictions. These findings represent associations with recorded business status, not causal effects or direct estimates of survival over time. Future research using longitudinal records and verified opening and closure dates could extend this framework to formal survival analysis.Keywords: Business Status Classification; Business Survival; Machine Learning; LightGBM; SHAP; Spatial Analysis; Business Clustering; San Diego County.