图结构特征挖掘在预测交易风险中的应用

Transaction Risk Prediction Using Graph Structure Feature Mining

交易风险预测是近年来电子商务和互联网金融领域关心的一个话题。传统的风控方法是基于具体规则来制定的,不能满足大数据时代应用的需要。比较流行的方法是先基于统计的特征挖掘,再基于特征进行模型训练,这也是比较传统的机器学习方法的工作模式。本文将改变传统的人工设计数据特征维度的方法,从图结构中自动地挖掘出特征,再结合最新机器学习Light GBM框架中的分类模型,进行用户交易风险分析。比传统的利用黑名单控制、基于统计特征分析的方式效果更好。同时,结合图的结构特征挖掘的方式在标签样本稀疏的情况下,效果也比传统方式更好。

Transaction risk prediction is a hot topic in Internet Banking and E-Commerce in recent years. Traditional risk control has to use empirical rules, which can not meet the needs of big data era applications. The conventional manner is to design statistical features engineering first of all, and then train a predictive model on the features. In this paper, we prepose to use the well-formed graph to automatically mine features from the graph structure, and then utilize the traditional machine regression model to analyze the users’ transaction risk. Compared with the traditional blacklist control methods and the statistical feature analysis methods, our proposed method is more effective. In addtion, as our method extracting from a wellformed graph, it outerperforms the tranditional risk prediction methods in the case of sparse samples.