To　address　dynamic　and　accurate　evaluation　of　driving　style　in　taxi　driver　safety　management,　this　paper　establishes　a　dynamic　recognition　model　of　driving　style　using　a　combination　of　unsupervised　clustering　and　supervised　classification.　Based　on　the　natural　driving　data　of　124　taxis　in　Beijing　for　one　month,　the　concept　of　vehicle　operation　information　entropy　is　proposed.　The　driver’s　safety　style　is　clustered　by　the　K-means++　clustering　algorithm　to　obtain　three　driving　styles:　“cautious,”　“aggressive,”　and　“normal.”　The　dynamic　recognition　model　of　driving　style　is　established　using　Gradient　Boosting　Decision　Tree　(GBDT),　support　vector　machine　(SVM),　and　logistic　regression　(LR),　and　the　effects　of　models　are　evaluated　and　compared.　Results　show　that　the　GBDT　algorithm　has　a　better　classification　effect　and　stronger　applicability　to　low-dimensional　data.　This　model　accurately　identifies　aggressive　drivers.　The　research　results　provide　support　for　drivers’　safety　management　and　targeted　intervention　in　the　taxi　industry.　©　ASCE.