《7-5 图机器学习在蚂蚁安全风控场景的应用.pdf》由会员分享,可在线阅读,更多相关《7-5 图机器学习在蚂蚁安全风控场景的应用.pdf(19页珍藏版)》请在三个皮匠报告上搜索。
1、图机器学习在安全风控的应用大安全机器智能朱 亮|01背景介绍02架构简介03安全风控图模型目录 CONTENT有向动态异质资金图主网络介绍DDGCL04展望背景介绍01|?$99.8$1000$199.8$299.8$301.9$2000$3000|背景介绍1.黑产会通过资金交易关系关联起来;2.资金流具有异常pattern;架构简介02|u1u2?Tree modelScoregraph_embse1ekfeaturesf1fmu1u2?u1u4u2u1u6u8u7u3u5?u24u10u11u21u18u22u23u16u17u19u20u14u13u12u15?u1u4u2u2u6u8u
2、7u3u5?u24u10u11u21u18u22u23u16u17u19u20u14u13u12u15?相较于树模型,图模型补充了两部分信息:k度子图内点边特征的聚合信息;k度子图内的拓扑结构信息;|算法架构简介安全风控图模型有向动态异质资金图主网络介绍DDGCL03|安全风控图模型有向动态异质图?点:账号、卡号、商家等;边:有方向,多种类型边关系;|安全风控图模型HGT将边方向作为一个边上一个属性|安全风控图模型HGT+DADEdge(Directional Attention Dual Edge Embedding)将边方向作为一个边上一个属性|安全风控图模型HGT+DADEdge(Dir
3、ectional Attention Dual Edge Embedding)/|安全风控图模型AUC提升千一打扰下召回提升TGAT00HGT+0.0327+5.634%HGT+DADEdge+0.0335+10.856%以某一个*场景为例,各模型效果对比如下:|安全风控图模型DDGCL(Debiased Dynamic Graph Contrastive Learning,CIKM2021)Many real-world graphs are dynamic in the sense that they evolve over time.Node vs representation will
4、 depend on its structural and compositional information,as well as the temporal information,and its representation shall be time dependent.Some methods derived from static graph scenarios are not directly applicable and may even lead to a questionable inference on these dynamic data.|安全风控图模型DDGCL假设:
5、大多数节点k度邻域子图在短时间内变化具有一致性Event stream!#=!%#!()Time-ware(GNN)EncoderTime-ware(GNN)EncoderTimeMaximizeTime-dependentAgreement)#!()#()%#()src dstFigure:Building blocks of DDGCL:As a contrastive-based self-supervisedframework,therere three ingredients of DDGCL:a Time-Aware GNN encoder,apositive sample reconstruction method,and a novel contrastive loss function.|安全风控图模型DDGCLDGI(Deep Graph Infomax)Loss:DDGCL Loss:|安全风控图模型DDGCL稳定模型训练,验证集过拟合缓解1%欺诈头部覆盖2.3%反洗钱头部覆盖2.9%向量召回新案件评估AUC展望04|展望 资金链表征进一步学习:MaskGAE、AdaPath;图模型鲁棒性的进一步提升:I.子图去噪、子图预计算;II.子图对抗攻击防御;III.DRO;图结构的进一步挖掘;非常感谢您的观看|