1、基于图神经网络的互联网金融欺诈检测敖 翔中科院计算所副研究员2022-05-21|01后疫情时代的后疫情时代的互联网金融欺诈互联网金融欺诈02用图神经网络用图神经网络为什么行？为什么行？03用图神经网络用图神经网络有什么坑？有什么坑？04用图神经网络用图神经网络有什么趋势？有什么趋势？目录目录CONTENT|后疫情时代的互联网金融欺诈01|金融欺诈威胁进一步增加FIS reports,the dollar volume of attempted fraudulent transactions rose 35%in April,2020,from a year earlier.Across al

2、l financial products,fraud rates rose by 33%in April 2020,compared with previous monthly averages.新冠疫情的爆发以前所未有的方式冲击了世界经济，进一步增加了遭受金融欺诈的风险。Fraud rate rises 33%during Covid-19 lockdown.https:/ fraud attempts rise during the coronavirus crisis.https:/ Products and Services如传统银行、保险、证券业务如电商平台、生活服务平台突出金融业务

3、远程银行（Remote banking）难以获得全面的客户身份验证信息，导致信用欺诈频发。阿里巴巴开始在其平台上向小企业主提供低息贷款，如生意贷。优步（Uber）在新冠第一波大流行期间为世界各地的弱势社区提供了 1000 万次免费乘车、用餐和送货服务。传统欺诈检测任务的挑战|类别不平衡概念漂移数据不可信传统欺诈检测Trustworthy?互联网欺诈检测任务的挑战|类别极度不平衡对抗攻击标注稀缺互联网欺诈检测互联网欺诈检测任务的挑战|类别极度不平衡对抗攻击标注稀缺互联网欺诈检测特征发现难互联网欺诈检测任务的挑战|类别极度不平衡对抗攻击标注稀缺互联网欺诈检测特征发现难样本价值敏感分布外样本特征学习

4、的指导信号弱用图神经网络为什么行？02|欺诈检测数据的演化|结构化数据半结构化数据非结构化数据量化表格访谈表格XML文件文本声音视频遥感数据欺诈检测模型所使用的数据类型和案例X.Zhu and X.Ao et al.,“Intelligent Financial Fraud Detection Practices in Post-Pandemic Era.”The Innovation,2021,2(4):100176.欺诈检测方法的发展|Rule-based systemsTraditional machine learningDeep learning1980s1990s2010s图神经网

5、络：一种新的趋势|User Profile:结构化、静态、数据缺失、噪声高User Behavior:时序、动态、高频User Relation:非欧氏数据异质图表征udvbca图神经网络方法发现欺诈用户、行为等活动新趋势：将多源和异构数据转换为（异质）图表示，并设计基于 GNN 的方法来发现欺诈活动特征学习能力、半监督学习多源异构数据的整合基于元路径的特征采样优化|Qiwei Zhong,Yang Liu,Xiang Ao*,Binbin Hu,Jinghua Feng,Jiayu Tang*,Qing He.Financial Defaulter Detection on Online C

6、redit Payment via Multi-view Attributed Heterogeneous Information Network.In WWW page 785-795,2020.(CCF A)问题：针对欺诈用户发现有效特征提取难贡献：将风控规则（知识）以元路径形式指导节点特征采样，优化特征提取效果：线上测试，逾期用户识别召回提升10.19%不同场景的风控规则表示为元路径同事朋友亲属BobMikeLily商户商户A A商户商户B B交易转账登录登录登录登录基于元路径的特征采样优化|Qiwei Zhong,Yang Liu,Xiang Ao*,Binbin Hu,Jinghua

7、 Feng,Jiayu Tang*,Qing He.Financial Defaulter Detection on Online Credit Payment via Multi-view Attributed Heterogeneous Information Network.In WWW page 785-795,2020.(CCF A)问题：针对欺诈用户发现有效特征提取难贡献：将风控规则（知识）以元路径形式指导节点特征采样，优化特征提取效果：线上测试，逾期用户识别召回提升10.19%同事朋友亲属BobMikeLily商户商户A A商户商户B B交易转账登录登录登录登录对于节点，根据元路

8、径采样的路径集合Paths based onMeta-pathPaths based onMeta-pathPaths based onMeta-pathPoolingPoolingPoolingMeta-pathAttentionNodeAttentionLinkAttentionConcat路径编码器元路径重要性节点和关系属性实验结果|Qiwei Zhong,Yang Liu,Xiang Ao*,Binbin Hu,Jinghua Feng,Jiayu Tang*,Qing He.Financial Defaulter Detection on Online Credit Payment

9、via Multi-view Attributed Heterogeneous Information Network.In WWW page 785-795,2020.(CCF A)数据：阿里巴巴真实数据 用户数量训练集：1.38 million(2019/01/01 2019/03/31)测试集：0.51 million(2019/05/01 2019/05/31）与当时最好方法 HACUDAAAI2019 相比，AUC指标提升2.3%，RP0.1指标提升7.4%与当时阿里巴巴线上的方法GBDT相比，AUC指标提升6.2%，RP0.1指标提升16.1%欺诈用户数量欺诈用户比例Ours用图神

10、经网络有什么坑？03|图神经网络核心思想：消息传递对于一个目标节点，根据周围的邻居通过消息传递生成embedding表示GCN、GraphSAGEGAT但是，我们面临的是类别不平衡图对于一个目标节点，根据周围的邻居通过消息传递生成embedding表示但是，我们面临的是类别不平衡图对于一个目标节点，根据周围的邻居通过消息传递生成embedding表示PC-GNN:面向类别不平衡图的采样GNN挑战：由于欺诈用户占比低，不利于图神经网络GNN消息传递贡献：改造GNN近邻采样机制，缓解类别不平衡问题效果：提升欺诈用户检测AUC 2.6%3.6%Yang Liu,Xiang Ao*,Zidi Qin,

11、Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)ewudgfvhbcaewudgfvhbcaewudvbcaewudfvbadvauw Pick-1 Choose-1 Pick-2Choose-2bvaucPick:根据标签类别分布占比，进行全局采样Choose:设计自适应的距离判别函数，进行局部结构优化PC-GNN:全局平衡采样Yang

12、 Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)Pick:根据标签类别分布占比，进行全局采样LF()=3LF()=6ewudgfvbcaudvbcaLabel FrequencySampling ProbabilityPC-GNN:局部结构调整Yang Liu,Xiang Ao*,Zidi

13、Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)Choose:对少数类节点实施自适应过采样，对所有节点实施自适应降采样少数类别过采样：所有类别降采样：对于少数类别:对于多数类别:udvbcaudvbcabvaucPC-GNN:整体网络结构Yang Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghu

14、a Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)ewudgfvhbcaewudgfvhbcaewudvbcaewudfvbadvauw Pick-1 Choose-1 Pick-2Choose-2bvaucRelation-1Relation-2FraudBenignPickChooseAggregatePC-GNN模型训练Yang Liu,Xiang Ao*,Zidi Qi

15、n,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)PC-GNN损失函数距离函数训练GNN网络参数训练损失函数实验:数据集Yang Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced

16、 Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)公开测试集-评论欺诈检测数据集YelpChi:Yelp评论（Hotel&Restaurant）Amazon:Amazon商品评论（Musical Instrument）真实数据集-阿里巴巴真实数据M7：2018/07/01-2018/07/31用户数据M9：2018/09/01-2018/09/30用户数据 Train/Valid/Test:40%/20%/40%实验：对比方法和评价指标Yang Liu,Xiang Ao*,Zidi Qin,Jian

17、feng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)对比方法GCN,GAT:传统GNNs方法DR-GCN:对偶正则GCN缓解类别不平衡GraphSAGE,GraphSAINT:基于采样的GNNsGraphConsis,CARE-GNN:基于图的欺诈检测SOTAPC-GNNP,PC-GNNC:PC-GNN的消融变体评价指标F1-macro:macro

18、average of F1-score of each classAUC:Area Under the ROC CurveGMean:Geometric Mean of True Positive Rate(TPR)and True Negative Rate(TNR)实验结果Yang Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3

19、168-3177,2021.(CCF A)RQ1:PC-GNN是否可以战胜已有的SOTA方法?与CARE-GNNCIKM20比AUC improvement 3.6%5.2%GMean improvement 0.6%3.7%实验结果Yang Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)R

20、Q2:消融实验，PC-GNN的各个模块对预测的贡献?Pick:全局采样是基础Choose:更进一步提升得益于局部结构优化实验结果Yang Liu,Xiang Ao*,Zidi Qin,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.Pick and Choose:A GNN-based Imbalanced Learning Approach for Fraud Detection.In WWW,page 3168-3177,2021.(CCF A)RQ3:参数敏感性实验 RQ4:Pick和Choose模块与传统GNN结合的效果AO-GNN:面向类别不平衡

21、图的AUC最大化GNN|优化AUC 的训练倾向于获得一个既能区分良性节点又能区分欺诈节点的模型Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)转化为鞍点搜索问题欺诈者可能主动伪装导致图结构已被“污染”|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi

22、,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)“污染”的图结构：欺诈节点经常通过与其他节点交互来混淆他们的身份欺诈者可能主动伪装导致图结构已被“污染”|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud D

23、etection.In WWW,page 1311-1321,2022.(CCF A)“污染”的图结构：欺诈节点经常通过与其他节点交互来混淆他们的身份是一个拓扑结构优化器结构的优化会带来预测结果的改变|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)预测结果的改变会引起AUC的变化|Mengda Huan

24、g,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)思路：向着AUC增大的方向优化图拓扑结构拓扑结构优化策略|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Ne

25、twork for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)拓扑结构优化策略：Environment:用环境GNN编码图的表示Action:“剪断”or“不剪”某一条边Reward:剪边后的图用GNN分类器预测得到的AUC-ROC变化值GNN分类器|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,p

26、age 1311-1321,2022.(CCF A)GNN分类器训练：1.用剪边策略对图结构进行“净化”2.在“净化”的图上，训练GNN分类器的参数，训练的Loss为AUC lossAO-GNN网络结构|Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)实验结果|公开测试数据集YelpChi:Yelp欺诈评

27、论Amazon:Amazon商品欺诈评论Books:图书虚假订单数据集Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)实验结果|RQ:AUC变化曲线RQ:何种类型的边被剪掉最多RQ：强化学习的效率验证Mengda Huang,Yang Liu,Xiang Ao*,Kuan Li,Jianfeng Chi,

28、Jinghua Feng,Hao Yang,Qing He.AUC-oriented Graph Neural Network for Fraud Detection.In WWW,page 1311-1321,2022.(CCF A)用图神经网络有什么趋势04|未来趋势如何防御潜在的对抗攻击？KDD2022 ,“场景依赖”,|“对抗攻击与防御”如何在新场景中快速自适应地学习特征？DASFAA2022如何更好利用无监督数据？KDD2022|,“预训练模型”|总结|1.互联网金融欺诈检测，用图神经网络为什么行？多源异构数据的整合天然兼容半监督学习场景2.互联网金融欺诈检测，用图神经网络有什么坑？消息传递机制过平滑不利于类别不平衡学习监督信号缺乏不利于有效特征提取3.互联网金融欺诈检测，还有什么趋势？场景依赖下的高效特征提取欺诈行为动态对抗建模大规模无监督行为数据的有效利用非常感谢您的观看|