Advanced fraud detection systems leverage the digital traces from (credit-card) transactions to detect fraudulent activity in future transactions. Recent research in fraud detection has focused primarily on data analytics combined with manual feature engineering, which is tedious, expensive and requires considerable domain expertise. Furthermore, transactions are often examined in isolation, disregarding the interconnection that exists between them.

In this paper, we propose CATCHM, a novel network-based credit card fraud detection method based on representation learning (RL). Through innovative network design, an efficient inductive pooling operator, and careful downstream classifier configuration, we show how network RL can benefit fraud detection by avoiding manual feature engineering and explicitly considering the relational structure of transactions. Extensive empirical evaluation on a real-life credit card dataset shows that CATCHM outperforms state-of-the-art methods, thereby illustrating the practical relevance of this approach for industry.

先进的欺诈检测系统利用来自（信用卡）交易的数字痕迹来检测未来交易中的欺诈活动。最近对欺诈检测的研究主要集中在数据分析与手动特征工程的结合上，这很乏味、昂贵并且需要大量的领域专业知识。此外，交易通常被孤立地检查，而忽略它们之间存在的互连。

在本文中，我们提出了CATCHM，一种基于表示学习 (RL) 的新型基于网络的信用卡欺诈检测方法。通过创新的网络设计、高效的归纳池算子和仔细的下游分类器配置，我们展示了网络 RL 如何通过避免手动特征工程和明确考虑交易的关系结构来有利于欺诈检测。对真实信用卡数据集的广泛实证评估表明，CATCHM优于最先进的方法，从而说明了这种方法与行业的实际相关性。