一种基于图卷积神经网络的网络小贷反欺诈方案

AN ANTI-FRAUD SCHEME BASED ON GRAPH CONVOLUTIONAL NETWORK FOR NETWORK MICRO-FINANCE

针对网络小额贷款团伙欺诈的现象,提出一种基于图卷积神经网络的网络小贷反欺诈方案。根据信贷领域知识自顶向下定义知识图谱,通过电话联系等数据搭建用户关系网络;对数据进行预处理,获取欺诈风险特征;将关系网络的邻接矩阵及特征作为图卷积网络的输入,聚合二阶邻居的特征传播计算未知标签节点的违约概率。实验表明,与DeepWalk模型结合逻辑回归、XGBoost、GBDT分类器相比,该方法在KS值上分别提高0.214、0.168、0.076,提高了正负样本区分度,能够有效识别团伙欺诈。

Aiming at the phenomenon of network microfinance gang fraud,this paper proposes an antifraud scheme based on graph convolutional network for network microfinance.According to the knowledge of credit field,the knowledge graph was defined by the topdown method,and the user network was built by the user call logs.The data was preprocessed to obtain the features of fraud risk.The adjacency matrix and features of the relational network were used as the input of graph convolutional network,and the feature propagation of the secondorder neighbors was aggregated to calculate the default probability of the unlabeled node.Experimental results show that compared with the DeepWalk model combined with logistic regression,XGBoost,GBDT classifier,the KS value of our method is improved by 0.214,0.168,0.076 respectively,which distinguishes between positive and negative samples better,and effectively identifies fraud gang.