基于特征和关联关系的社交平台欺诈检测

Social platform fraud detection based on association and usercharacteristics

各个社交平台的作弊问题日趋严重,欺诈检测工作越来越有必要.现有在该场景的解决办法没有同时利用用户特征和关联关系两方面重要信息或者不能应用于现实上亿规模的数据量.针对这个问题,开创性地将GraphSAGE算法应用于社交平台的反作弊场景并进行改进,提出带权采样GraphSAGE算法.改进后算法根据节点之间特征相似程度进行采样.在真实大数据集上进行了实验,线下实验中,相较于基准模型和现有主流模型,性能上有了较明显的提升,且加快了模型的收敛过程.在线上结合基础规则,达到了极高的精确率,并召回之前未察觉的两个作弊团伙.

The problem ofcheating on various social platforms is getting more serious and fraud detection is becoming more and more necessary. The existing solutions in this scenario have the problem of not using both important information of user characteristics and association relationships at the same time, or cannot be applied to large-scale datasets in reality. Aiming at this problem, this paper attempts the application of improved GraphSAGE algorithm to the anti-cheating scenarios on the social platform, the weighted GraphSAGE algorithm is proposed which sampling based on the degree of feature similarity between nodes. The experiments are performed on large-scale real-world datasets. In the offline experiment, the fraud detection performance is significantly improved compared with the benchmark model and the existing mainstream model. In addition, the convergence process of the model is accelerated. Combining the basic rules in the online, it can achieve high accuracy and recalled two fraud groups undetected before.