Detecting Anomalous line-items by Modeling the Legal Case Lifecycle
share
Anomaly detection continues to be the subject of research and development efforts due to its wide-ranging applications and the value of detecting anomalous patterns in a variety of use cases. While many anomalies may be relatively benign, others may be more severe have the potential to significantly impact to the business, user, or other involved parties. In finance, detecting anomalous transactions can provide dramatic improvements to financial audits, given that many audits continue to involve significant human effort in review of accounting documents. In the case of the legal industry - which is the focus of this work - detecting anomalies is important to both data and legal integrity, and serves a secondary function in supporting downstream analytics and machine-learning capabilities. In this work, we detail an approach for detecting anomalous activities based their suitability in the legal case's lifecycle (modeled using a set of case-level and invoice line-item-level features). We illustrate our approach for invoice line-item anomaly detection which works in the absence of labeled data, by utilizing a combination of subject matter expertise and synthetic data generation for model training. We characterize the method's performance using a set of well understood, easily accessible model architectures. We demonstrate how this process provides a path towards solving certain anomaly detection problems when the characteristics of the anomalies are well known, and offer lessons learned from applying our approach to detecting potentially anomalous line-items on real-world data.
READ FULL TEXT
