To　advance　the　interpretability　of　machine　learning　for　long-term　crime　prediction　in　China,　we　compared　the　performance　of　multiple　machine　learning　algorithms　in　predicting　the　spatial　pattern　of　theft　in　Beijing.　Gradient　boosting　decision　tree　emerged　as　the　algorithm　with　best　predictive　accuracy.　After　identifying　the　importance　of　criminogenic　features,　we　extended　the　interpreter　SHAP　to　reveal　nonlinear　and　spatially　heterogeneous　associations　between　environmental　features　and　theft　and　we　summarized　six　relation　types　of　such　associations　at　the　global　scale.　At　the　local　scale,　we　clustered　six　area　types　according　to　the　contribution　of　environmental　attributes　to　theft　prediction　in　each　grid.　Policy　makers　should　adopt　place-based　crime　prevention　measures　based　on　the　specific　type　of　each　grid　belongs　to.