BREAD Offers Scalable, Accurate Anomaly Diagnosis for AI Systems

As AI-based anomaly detection systems become more prevalent in complex, high-dimensional environments, the challenge of diagnosing *why* an anomaly occurred intensifies. Traditional diagnosis methods are often tied to specific detection models and don't translate well to advanced AI techniques like statistical process monitoring (SPM). While model-agnostic explainable AI (XAI) offers a general framework, existing methods frequently suffer from scalability issues or incorrectly attribute relevance to noise features, compromising diagnostic accuracy. To overcome these limitations, researchers propose BREAD (Baseline-Referenced Explanations for Anomaly Diagnosis). This new method is designed to be scalable and leverages both the anomalous observation and crucial information from a normal baseline. By incorporating this dual perspective, BREAD can more accurately pinpoint the features responsible for an anomaly. Mathematical guarantees provided in the paper show that BREAD achieves higher faithfulness in detecting anomaly-causing features, particularly in mean-shift anomaly scenarios, compared to methods like LIME. Both simulation studies and a real-world case study validate BREAD's effectiveness, demonstrating its ability to generate more faithful and accurate diagnostic results for AI-based prospective anomaly detection.