With　the　development　of　information　and　communication　technology,　the　situation　of　communication　frauds　is　becoming　more　and　more　serious,　how　to　identify　fraudulent　telephone　accurately　and　effectively　has　become　an　urgent　task　in　telecom　operation　at　present.　Affected　by　the　power　law　distribution,　existing　machine　learning　methods　are　used　to　identify　the　unbalanced　distribution　data　set　of　positive　and　negative　samples　with　low　recognition　accuracy.　This　paper　proposes　ADASYN+RF　model.　First　of　all,　for　the　problem　of　unbalanced　data　sets,　this　paper　chooses　the　ADASYN(Adaptive　Synthetic　Sampling)　algorithm　to　rebalance　the　original　data　set.　Secondly,　we　choose　the　random　forest　algorithm　is　employed　to　train　the　new　data　set　to　avoid　overfitting.　Finally,　two　groups　of　comparative　experiments　are　carried　out　respectively,　and　the　results　show　that:　(1)　For　the　processing　of　biased　data,　the　ADASYN　algorithm　used　in　this　paper　is　more　advantageous　than　the　traditional　SMOTE　algorithm;(2)　Compared　with　Non-integrated　learning　model,　the　accuracy,　recall　rate　and　F1　value　of　the　ADASYN+RF　model　are　significantly　improved.