Recent　years　witnesses　the　rampancy　of　telephone　fraud　along　with　the　development　of　modern　communication　technology.　The　challenges　from　telephone　fraud　identification　mainly　exist　in　two　aspects:　(1)　the　telephone　fraud　records　are　typical　imbalanced　data　due　to　the　characteristic　of　heterogeneous　spatial-temporal　distribution,　leading　to　bias　towards　predicting　the　majority　class;　(2)　traditional　evaluation　metrics　in　imbalanced　learning　mainly　rely　on　accuracy　or　precision,　neglecting　the　completeness　of　telephone　fraud　identification　in　real-world　implementations.　In　response　to　the　limitations　of　traditional　methods,　we　propose　the　Stacked-SVM　framework　based　on　heterogeneous　ensemble　learning　and　support　vector　machines　(SVMs).　We　first　employ　both　edited　nearest　neighbors　(ENN)　and　adaptive　synthetic　sampling　(ADASYN)　to　alleviate　the　high　dimensional　curse　in　imbalanced　data　resampling;　secondly,　we　propose　the　optimal　linear　combination　strategy　in　the　iteration　of　Stacked-SVM　and　demonstrate　its　validity　with　the　help　of　Kullback-Leibler　divergence.　Finally,　we　construct　the　Stacked-SVM　framework　with　respect　to　the　constraints　of　the　loss　function　in　SVM.　We　further　compare　the　performance　under　different　evaluation　metrics　(i.e.,　accuracy,　precision,　recall,　F1-score,　and　AUC　value)　with　other　four　traditional　telephone　fraud　identification　methods,　namely　Logistic　Regression,　Isolation　Forest,　SVM　with　random　parameter　settings,　and　optimized　SVM.　We　implement　Stacked-SVM　with　a　list　of　experiments　based　on　real　telephone　fraud　data　sets　in　the　form　of　calling　detail　records　(CDRs)　from　a　Chinese　domestic　telecom　operator.　The　experimental　results　show　that　the　proposed　Stacked-SVM　holds　a　93.83%　recall　value　and　an　82.96%　accuracy　in　telephone　fraud　identification,　behaving　more　precise　and　robust　than　other　models.　©　2019　ACM.