The　prevention　and　crackdown　of　fraud　calls　have　been　paid　more　and　more　attention　by　industrial　and　academic　societies.　Most　current　researches　based　on　machine　learning　ignore　the　imbalanced　data　distribution　characteristic　between　normal　and　fraudulent　call　users,　and　the　outputs　neglect　the　probability　fluctuation　range　of　the　suspected　fraudulent　calls.　To　overcome　these　limitations,　we　first　construct　user　behavioral　feature　vector　by　a　random　forest　method.　Secondly,　we　propose　a　novel　hierarchical　sampling　method　to　overcome　the　class　imbalance　problem.　Thirdly,　we　propose　a　novel　fraud　call　recognition　method　based　on　HPO-LGBM　(the　Bayesian　hyper　parameter　optimization　based　on　random　forest　and　Light　Gradient　Boosting　Machine).　Finally,　we　further　evaluate　the　method’s　performance　with　a　DRI　(dynamic　recognition　interval)　model.　Experimental　results　on　public　datasets　show　that　the　proposed　HPO-LGBM　holds　a　92.90%　F1　value,　a　91.90%　AUC,　a　92.92%　G-means,　and　a　92.37%　MCC　in　fraud　call　recognition.　In　addition,　the　proposed　HPO-LGBM　model　can　further　give　the　dynamic　recognition　interval　of　the　output　result,　behaving　more　robust　than　other　models　(i.e.,　LR,　RF,　MLP,　GBDT,　XGBOOST,　LGBM).　©　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.　2024.