Unsupervised Anomaly Detection for Financial Auditing with Model-Agnostic Explanations

Explainable Artificial Intelligence (AI) has emerged to be a key component for Black-Box Machine Learning (ML) approaches in domains with a high demand for transparency. Besides medical expert systems, which inherently need to be interpretable, transparent, and comprehensible as they deal with life-changing decision tasks, other application domains like financial auditing require trust in ML as well. The European General Data Protection Regulation (GDPR) also applies to such highly regulated areas where an auditor evaluates financial transactions and statements of a business. In this paper we propose an ML architecture that shall help financial auditors by transparently detecting anomalous datapoints in the absence of ground truth. While most of the time Anomaly Detection (AD) is performed in a supervised manner, where model-agnostic explainers can be easily applied, unsupervised AD is hardly comprehensible especially across different algorithms. In this work we investigate how to dissolve this: We describe an integrated architecture for unsupervised AD that identifies outliers at different levels of granularity using an ensemble of independent algorithms. Furthermore, we show how model-agnostic explanations can be generated for such an ensemble using supervised approximation and Local Interpretable Model-Agnostic Explanations (LIME). Additionally, we propose techniques for explanation-post-processing that allow explanations to be selective, receiver-dependent, and easily understandable. In a nutshell, our architecture paves the way for model-agnostic explainability for the task of unsupervised AD. It can further be transferred smoothly to other unsupervised ML problems like clustering problems.