Explainable AI Frameworks for Enterprise Reporting Anomaly Attribution

Abstract

In enterprise financial reporting environments, there are increasingly sensitive anomaly detection systems in place, with algorithms that can detect deviations in statistics very accurately, but the identification of such deviations to meaningful causal features is relatively opaque to the business and the regulatory auditor. This paper summarizes all of the key knowledge on the Explainable Artificial Intelligence (XAI) frameworks tailored for enterprise reporting anomaly attribution, from anomaly detection basics to current interpretability techniques published up to 2020. The proposed architecture combines density-based local outlier detection [3] with tree-ensemble isolation methods [11] and deep neural network autoencoders [7] as well as post-hoc attribution layers, such as SHapley Additive exPlanations [13], Local Interpretable Model-agnostic Explanations [16], Layer-wise Relevance Propagation [18] and counterfactual explanation generators that comply with GDPR Article 22 requirements [20]. Experimental results show that the accuracy of SHAP based attribution is 91.4% and 88.3% scalable with 1 million records, while the comprehensibility of LIME in user studies (n=45) is 91%. For baselines with more than one million records, this model achieves an AUC-ROC of 92.1%. On datasets of over 1 million records, this model achieves an AUC-ROC of 92.1%. Regulatory compliance scores of 95.0% are achieved under GDPR evaluation protocols. The synthesis shows that there is no one-size fits all XAI method—which motivates the development of a hybrid attribution stack that is detailed herein—as it relates to enterprise use cases. Findings place the proposed framework in the context of an organisation that needs to have a transparent and auditable financial reporting pipeline to provide anomaly attribution.