The　nonlinear　relationship　and　spatially　heterogeneous　relationship　between　environmental　factors　and　criminal　activities　are　the　main　reasons　for　both　the　theoretical　and　empirical　divergence,　but　the　relevant　analysis　remains　fragmented　and　faces　limitations　such　as　linear　relationship　hypothesis,　collinearity　problems　and　omitted　variable　bias.　This　study　uses　Gradient　Boosting　Decision　Tree　(GBDT)　algorithm　and　Shapley　Additive　Explanation　(SHAP)　interpreter　in　machine　learning　to　systematically　reveal　the　nonlinear　and　spatially　heterogeneous　relationships　between　48　built　and　social　environmental　factors　on　violent　crime　in　Beijing.　Our　research　has　revealed　the　existence　of　seven　distinct　types　of　nonlinear　relationships　between　environmental　factors　and　violent　crime,　each　exhibiting　unique　trends　in　the　direction　of　influence　and　marginal　effects.　Furthermore,　we　have　found　that　the　association　between　environmental　factors　and　violent　crime　exhibits　varying　degrees　of　spatial　heterogeneity.　By　utilizing　K-means　clustering　analysis,　the　entire　area　can　be　segmented　into　six　distinct　regions,　each　characterized　by　different　critical　criminogenic　factors.　These　findings　suggest　that　the　applicability　of　crime　geography　theories,　such　as　the　classification　of　crime　generators,　attractors,　and　inhibitors　based　on　crime　pattern　theory,　the　validity　of　street　eye　theory　and　defensible　space　theory,　and　the　impact　of　social　attributes　as　proposed　by　social　disorganization　theory,　may　depend　on　the　value　range　of　environmental　factors　and　differ　across　locations.　In　light　of　these　findings,　it　is　recommended　that　crime　prevention　strategies　shift　from　universal　to　targeted　approaches,　wherein　public　resources　are　allocated　to　specific　value　ranges　of　environmental　variables　and　prioritized　regions.　©　2024　Science　Press.　All　rights　reserved.