Core Concepts
AcME-AD offers efficient interpretability for anomaly detection models, enhancing trust and usability in practical scenarios.
Abstract
AcME-AD introduces a novel approach rooted in Explainable Artificial Intelligence principles to clarify Anomaly Detection models. It provides local feature importance scores and a what-if analysis tool to aid root cause analysis and decision-making. The method transcends model-specific limitations by offering an efficient solution for interpretability, validated on synthetic and real datasets.
Traditional Anomaly Detection methods excel at identifying outliers but lack transparency, hindering their adoption in critical scenarios. AcME-AD addresses this gap by providing insights into the factors contributing to anomalies, enabling better decision-making. The approach is computationally efficient, making it suitable for time-critical applications like intrusion detection or fault detection.
Research interest in Explainable Artificial Intelligence (XAI) has recently shifted towards unsupervised tasks like Anomaly Detection. AcME-AD stands out by focusing on local interpretability, analyzing individual features' influence on anomaly predictions. The method's sub-scores offer valuable insights into feature importance and classification changes.
In experiments on synthetic and real-world datasets, AcME-AD demonstrates its effectiveness in explaining anomalies with high precision. Comparisons with KernelSHAP and LocalDIFFI show consistent feature rankings across different methods. Additionally, the method outperforms KernelSHAP in computational efficiency, making it ideal for rapid interpretability needs.
Feature selection experiments further validate the relevance of features identified by AcME-AD, showcasing improved model performance compared to random feature selection strategies.
Stats
Pursuing fast and robust interpretability in Anomaly Detection is crucial.
AcME-AD offers local feature importance scores and a what-if analysis tool.
The method transcends model-specific limitations by providing an efficient solution for interpretability.
AcME-AD demonstrates effectiveness with tests on both synthetic and real datasets.
The approach is computationally efficient, making it ideal for time-critical applications.