Abstract

In this article, we study sequences of events with unknown event types commonly arising in customer service processes, such as claim processing, and present a novel approach for detecting outliers and online monitoring of the progression of such sequences until termination. In our approach, we create a series of semi-parametric Cox PH regression models (called sub-models) pertaining to each event occurrence and then use their residuals to create a multivariate normal monitoring statistic. Owing to the benefits of Cox PH, the method can explicitly incorporate relationships to past events. Furthermore, the statistic is also adjusted for non-termination which makes it usable in the online setting. We demonstrate the efficacy of our method through a numerical study and application to a real-world data set obtained from a property and casualty insurance company.

摘要

在本文中，我们研究了客户服务流程（例如索赔处理）中常见的未知事件类型的事件序列，并提出了一种检测异常值和在线监控此类序列直至终止的进展的新方法。在我们的方法中，我们创建了一系列与每个事件发生相关的半参数 Cox PH 回归模型（称为子模型），然后使用它们的残差来创建多变量正常监测统计量。由于 Cox PH 的好处，该方法可以明确地合并与过去事件的关系。此外，统计数据还针对非终止进行了调整，使其可用于在线设置。