This article reports the results of our work to construct a system for the detection of fraudulent behavior in procurement transactions. To solve the problem, we model different types of fraud via separate statistical indicators. We propose a formalized framework to describe the severity of fraud in a unified way regardless of underlying fraud mechanics. Subsequently, we leverage this concept to build indicator ensembles that collect evidence from multiple indicators and deliver an interpretable per transaction score to the procurement audit officer. As a case study, we overview 48 such fraud indicators constructed for our client and describe two examples in detail showing how our formal definitions can be transformed into a practical implementation. The presented results include experiments with all indicators on data covering four years of procurement activity with approximately 216,000 transactions coming from a large government organization in Singapore. The final evaluation of our system shows 67.1% precision in detecting suspicious transactions. The article describes how outcome of our work helped to effectively cope with the problem of anomaly detection explainability and the lessons learned from integrating this solution to operational practices of a procurement department.

中文翻译：

---

采购欺诈识别的可解释异常检测——来自实际部署的经验教训

本文报告了我们构建用于检测采购交易中的欺诈行为的系统的工作结果。为了解决这个问题，我们通过单独的统计指标对不同类型的欺诈进行建模。我们提出了一个正式的框架，以统一的方式描述欺诈的严重程度，而不管潜在的欺诈机制如何。随后，我们利用这一概念构建指标集合，从多个指标收集证据，并向采购审计官提供可解释的每笔交易评分。作为案例研究，我们概述了为我们的客户构建的 48 个此类欺诈指标，并详细描述了两个示例，展示了如何将我们的正式定义转化为实际实施。所呈现的结果包括对涵盖四年采购活动数据的所有指标进行的实验，其中大约有 216,000 笔交易来自新加坡的一个大型政府组织。我们系统的最终评估显示，检测可疑交易的准确率为 67.1%。文章描述了我们的工作成果如何帮助有效应对异常检测可解释性问题，以及将此解决方案集成到采购部门的运营实践中的经验教训。