一种面向网络支付反欺诈的自动化特征工程方法  

An Automated Feature Engineering Method for Online Payment Fraud Detection

作　　者：王成[1,2,3] 王昌琪[1,2] WANG Cheng;WANG Chang-Qi(Department of Computer Science and Technology,School of Electronics and Information Engineering,Tongji University,Shanghai 201804;The Key Laboratory of Embedded System and Service Computing,Ministry of Education,Shanghai 201804;Shanghai Institute of Intelligent Science and Technology,Tongji University,Shanghai 200092)

机构地区：[1]同济大学电子与信息工程学院计算机科学与技术系,上海201804 [2]嵌入式系统与服务计算教育部重点实验室,上海201804 [3]同济大学上海智能科学与技术研究院,上海200092

出　　处：《计算机学报》2020年第10期1983-2001,共19页Chinese Journal of Computers

基　　金：国家自然科学基金(61972287);2018年上海市青年拔尖人才开发计划;同济大学研究生教育改革与研究项目(ZD1903031)资助

摘　　要：互联网金融欺诈正导致诸多社会经济问题.网络支付是互联网金融中的典型模式之一,此模式中的欺诈交易也是互联网金融欺诈的主要形式之一.通过构建基于机器学习的欺诈检测模型来识别欺诈交易的方法已成为网络支付反欺诈领域的主流思路.在构建欺诈检测模型的过程中,特征工程是最为关键的一步,特征的质量将直接影响模型的性能;通常,这也是最为耗时且对相关领域的专业知识要求最高的步骤.现有网络支付欺诈检测模型在特征工程上主要是领域专家基于业务知识以手动构造的形式来开展.而在网络支付模式下欺诈场景众多,不同场景下的特征构造流程不尽相同.人工特征构建方法已不能满足与日俱增的反欺诈需求.解决此问题的重要方法之一便是自动化特征工程.本文针对网络支付欺诈检测提出了一种轻量化、树结构、高效率、可扩展和可解释的自动化特征工程方法.该方法:(1)对计算条件的要求低且对数据集样本的依赖性小,这一优势是利用树结构模型进行特征构造得以实现;(2)可构造出深度层次的复杂特征和广度层次的各类型特征,这一优势是利用节点处特征构造的新型流程和转换函数权重向量的时效性更新机制得以实现;(3)在网络支付模式不同场景下可实现跨场景复用,这一优势是通过复用和扩展定制化转换函数得以实现;(4)构造出的特征具有可解释性,这一优势得益于基于结合转换函数与树模型的特征构造过程具备可表达性.本文在网络支付典型场景的业务数据集上验证了所设计的自动化特征工程方法的有效性.Internet finance fraud is an increasingly serious social and economic problem.Online payment services(OPSs)are the typical models of Internet finance,and the fraudulent transaction in OPSs is also a typical fraud pattern.The method of identifying fraudulent transactions by constructing a fraud detection model based on machine learning has become a promising idea for online payment anti-fraud.In the process of constructing fraud detection models,the feature engineering is the most critical step.It is also one of the most time-consuming and specialized steps in the relevant area.In the study of feature engineering,the existing online payment fraud detection models are mainly carried out by experts in the form of manual construction based on business knowledge.However,there are many fraud scenarios in OPSs where the process of feature construction is so different.Artificial feature construction methods can no longer meet the increasing demand of anti-fraud.An important way to solve this problem is to automate feature engineering.In the field of Internet financial anti-fraud,the expressibility and interpretability of features play a pivotal role.It is helpful to understand the original source fields and their construction process of important features.This is useful for mining and analyzing the characteristics of fraud methods and follow-up improvement rules engines.These are of great significance for fraud detection models.Therefore,the interpretability of the model method is particularly important.Usually,the optimization of detection accuracy is carried out under the premise of ensuring interpretability.This paper proposed a lightweight,tree-structure,high efficiency and scalable automatic feature engineering method for fraud detection of online payment.The method is as follows:(1)The method has low requirements on the calculation conditions and little dependence on the dataset samples.To realize this advantage,it used the tree structure model to construct the features.(2)The method can construct complex features

关 键 词：网络支付 互联网金融 欺诈检测 自动化特征工程 机器学习 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]