金融交易风险防控是金融风控平台最重要的能力之一.近年来,随着金融风控平台智能化需求的不断升级,对其中相关应用算法的性能要求也水涨船高.目前业界已完成了两代针对金融交易行为的表征学习框架的迭代升级.第一代框架引入了金融交易活动参与者自身的历史行为序列,利用序列模型学习其历史行为特征.第二代框架通过一套实时大数据系统对资金流图进行建模,根据业务专家预定义的业务规则计算出需要的实时特征,并将其输入到后续的判别模型中.相比第一代,第二代框架引入了更多实时动态资金流图上的交互信息,因而取得了不错的性能提升.然而,第二代框架在精细化、智能化和时序建模方面仍存在较大不足.为了解决这些问题,本文针对性地设计了第三代框架,该框架通过动态图表征学习算法,从实时资金流图的原始数据中直接进行表征学习,以此规避了第二代框架中的诸多问题.总的来说,本文在时序信息建模和动态图框架层面均进行了创新性设计.在时序信息建模层面,利用了 C2GAT模块(连续时间和上下文感知的图注意力神经网络),在动态多变的资金流图上快速地捕捉了高阶的结构化时序状态与信息.在动态图框架层面,开发了一套联合实时动态图表征框架——RULF,该框架可以实时刻画出金融场景中多用户资金行为中存在的特定模式.将金融场景中多角色联合行为和单角色独立行为进行了显式地解耦,并将多个子图模块联合起来学习,通过学习到更精准的行为表征,从而更进一步地提高下游判别模型的性能.本文将以花呗套现交易识别—一个典型的金融交易风控场景为例,介绍该框架在实际业务场景中的设计思想和实现细节.
In recent years,with the continuous escalation of demand in the intelligent financial platforms,the perfor-mance requirements for these relevant application algorithms in financial scenarios have also risen.At present,two genera-tions of frameworks about financial role representation learning have been widely used in the industry.The first-generation framework introduced the unique historical sequence of financial roles,and used the sequence model to learn the historical behavior of the role.The second-generation framework put more emphasis on the interaction between roles,built a real-time dynamic graph system through capital flow,and directly obtained the required real-time features through graph calcula-tion according to predefined business rules,and added them to the follow-up discriminant models.Compared with the first generation,it introduced more interactive information,resulting in a good performance improvement.However,the second-generation framework still has great limitations in terms of timeliness,generalization,and ease of use.In order to solve these problems,we design the third-generation framework which directly builds feature from the original real-time capital flow graph through the dynamic graph learning algorithm,avoiding many problems in the second generation.This paper mainly carries on the innovative design in temporal modeling and frame design.In terms of temporal modeling,we design the C2GAT to flexibly capture high-order structured temporal information on dynamic graphs.In terms of framework mod-eling,we design a real-time dynamic graph framework-RULF,which can better capture and characterize the specific pat-terns existing in capital behavior in real time financial scenarios.We explicitly separate multi-role joint behavior and single-role independent behavior in financial scenarios,and jointly learn multiple subgraph modules to obtain accurate user repre-sentation and performance improvement.A typical interactive financial scenario will be used as a credit cashback example in this article to introduce our design ideas and implementation details in actual business scenarios.