深度学习在信息融合和欺诈风险识别中的应用

整理文档很辛苦,赏杯茶钱您下走!

免费阅读已结束,点击下载阅读编辑剩下 ...

阅读已结束,您可以下载文档离线阅读编辑

资源描述

深度学习在信息融合和欺诈风险识别中的应用confidentialmaterialfromTencentCloud腾讯云业务安全中心•WEgraduatefrom:•NUS•UniversityofSouthCarolina•GeorgiaTech•UniversityofWashington•ImperialCollegeLondon•UniversityofWarwick•…•核心业务:•金融反欺诈•防刷•内容安全•…USTC•BachelorNUS•PhDMozat•HeadofDataIntelligencePropertyGuru•HeadofDataScienceTencentCloudbusinesssecuritycenter•LeadDataScientist•Aboutme:confidentialmaterialfromTencentCloud安全常识1.中国金融科技浪潮confidentialmaterialfromTencentCloudJourneytothe“普惠金融”2010:Consumerfinanceemerged2015:monthlyP2Plendingexceeds100Billion2017:PaipaiDai2017H1netprofitexceeds1billion2017Nov.:《关于立即暂停批设网络小额贷款公司的通知》2017Dec.:《关于规范整顿“现金贷”业务的通知》2018:BanksandinternetcompanyactivelyentertheonlinelendingmarketEIRmustnotbehigher36%ThemoneymustbeusedforSpecificPurpose2018Apr.:DiDilaunch滴水贷2018May.:Huaxialaunch龙商贷2.黑产confidentialmaterialfromTencentCloud互联网金融的发展催生了黑产的繁荣该网贷产品日均申请量超过10万件,日均放款额超过一亿元,如果没有风控,日均损失金额将超过2000万1.5Billionfromcollusion银行卡盗刷行为猖獗“医美骗贷大狂欢,中介医院勾结撸出15个亿”confidentialmaterialfromTencentCloud黑产-数据交易confidentialmaterialfromTencentCloud黑产-养号机器人confidentialmaterialfromTencentCloud黑产-扫村3.风控算法的演进confidentialmaterialfromTencentCloud策略系统•Pros•Definite•Fastdeployment•Nomath•Cons•Relayonexperience•DifficulttomaintainconfidentialmaterialfromTencentCloud评分卡LRiswidelyusedbecause•Interpretable•Goodgeneralization•Fast•SimpleLogisticRegressionconfidentialmaterialfromTencentCloud逻辑回归的高级玩法•Featureexplosion•Featurecrossing•Ensemblemodel•Featureevaluation•IV•Featureselection•Avoidmulti-collinearity=𝑃𝑦𝑖−𝑃𝑛𝑖∗ln𝑃𝑦𝑖𝑃𝑛𝑖𝑖=𝑦𝑖𝑦𝑠−𝑛𝑖𝑛𝑠ln⁡𝑦𝑖/𝑦𝑠𝑛𝑖/𝑛𝑠𝑖confidentialmaterialfromTencentCloudOne-hotencoding•[“fromEurope”,“fromUS”,“fromAsia”]•-[0,1,2]?•-[001,010,100]•[“male”,“female”]•-[01,10]•[“usesFirefox”,“usesChrome”,“usesSafari”,“usesInternetExplorer”]•-[0001,0010,0100,1000]•[“male”,“fromAsia”,“usesChrome”]•-[01,100,0010,]•[userid]-?confidentialmaterialfromTencentCloudXGBoost-简单好用性能优•Thego-to”choiceformachinelearningprojects•Ensemblemodel•Robust•Goodbalancebetweenaccuracyandgeneralizationsource:“好的架构是包容所有好技术的重要前提“confidentialmaterialfromTencentCloud每个神经元都是一个LRconfidentialmaterialfromTencentCloudPCA降维=AutoencoderLoss=1𝑁𝑥𝑖−𝑥𝑖2𝑁𝑖=1EncoderDecoderconfidentialmaterialfromTencentCloud编码器和查找表labelembeddedchrome100firefox0.510safari00.51ie010.5𝐸𝑗(𝑋𝜃)=𝑤𝑖𝑗𝑥𝑖𝑖,=𝑤𝜃𝑗,⁡𝑤ℎ𝑒𝑟𝑒⁡𝑥𝜃≠⁡0𝐸𝑗confidentialmaterialfromTencentCloudCollaborativefiltering=浅层神经网络P𝐼𝑚𝑈𝑛=𝑤𝑖𝑛𝐸𝑖𝑖=𝑤𝑖𝑛𝑤𝑚𝑖𝑖UserItem𝑤𝑖𝑛𝑤𝑚𝑖𝐸𝑖UserItemLookupEmbedding点乘Onehot求和confidentialmaterialfromTencentCloud集成学习=DropoutEachpersonhas2/3chancetopredictthegameresultcorrectly.Whatiftheyvote?𝐶32⁡∗232∗13+233=202723?confidentialmaterialfromTencentCloud风险预测流程UnderstandhowtheundergroundindustryworksCollectallthedatathatmightberelatedLabelyourdataDesignnetworkandtrainMonitoringtheperformanceandupdatethemodelconfidentialmaterialfromTencentCloud混合神经网络•略Q&AThanks!

1 / 26
下载文档,编辑使用

©2015-2020 m.777doc.com 三七文档.

备案号:鲁ICP备2024069028号-1 客服联系 QQ:2149211541

×
保存成功