BREAD: Baseline-Referenced Explanations for Anomaly Diagnosis

BREAD: Baseline-Referenced Explanations for Anomaly Diagnosis proposes a scalable, model-agnostic explanation method for AI-based prospective anomaly detection. The method addresses the challenge of diagnosing which features drive a flagged observation away from normal behavior, a task for which traditional SPM diagnosis methods are often model-specific and existing XAI methods like LIME suffer from scalability limitations or assign relevance to noise features.

The paper states: "We propose a scalable, baseline-referenced diagnosis method that uses both the anomalous observation and normal baseline information. We provide mathematical guarantees that under a mean-shift anomaly setting, the proposed method achieves higher faithfulness in detecting the features causing the anomaly compared to LIME."

The core idea is described as: "BREAD modifies the LIME algorithm by introducing a novel sampling scheme and a sampling-aligned weighting kernel. The main idea is to fit a regression model on the weighted sampled data around both the anomalous observation and a reference normal point. By introducing the baseline information into the surrogate construction, the resulting explanations become sparser, concentrating relevance on anomaly related variables while assigning near-zero relevance to irrelevant features."

The method works by generating synthetic samples around both the anomalous observation tj∗ and a reference normal point tref, then fitting a weighted ridge regression surrogate. The weighting kernel is defined as πi = exp(−d2tj∗,tref(zi)/σ2), where the distance is dtj∗,tref(zi) = √(∥zi − tj∗∥2 · ∥zi − tref∥2). This distance assigns large weights to samples that are close to either the detected anomaly or the reference normal point.

The paper provides theoretical guarantees, including Theorem 3.3 (Asymptotic sparsity): "Suppose the shift setup in (5) holds, the synthetic data matrix Z(δ) satisfies Assumption 3.1, and the corresponding outputs satisfy Assumption 3.2. Let β̂ denote the estimator in (4) and define ϵ = ∥β̂R∥2/∥β̂K∥2. Then, ϵ → 0, as δ → ∞." This establishes that as the anomaly moves further from the normal reference, relevance concentrates on the truly shifted features.

The paper also demonstrates that LIME lacks this guarantee: LIME does not enjoy this guarantee that the ratio of coefficients need not converge to zero, even as δ → ∞. In a two-dimensional example, LIME's coefficient ratio is bounded away from zero by a constant ϵL ≈ γ/2, regardless of how large δ becomes, which is a fundamental limitation that BREAD overcomes.

The simulation study evaluates the method across three dimensions: faithfulness, robustness, and stability. The paper reports: "BREAD achieves the highest faithfulness score, while LIME remains relatively low and flat. Moreover, the performance gap between these two methods increases rapidly as the shift becomes large, which is consistent with Theorem 3.3. For robustness, BREAD achieves the highest robustness for moderate-to-large mean shifts. For stability, BREAD achieves a consistently high faithfulness score even at a relatively small sample size, and its standard deviation decreases rapidly as the number of samples increases."

The real-world case study involves escalator monitoring with unexpected shutdowns. The paper states: "The BREAD method produces markedly different and more faithful results. Escalator 08 receives the largest negative importance (median ≈ −0.80), and Escalators 13 and 14 receive consistently large negative importance (median ≈ −0.25–−0.50). The remaining 13 escalators are assigned near-zero importance with low variance, reflecting the asymptotic sparsity property established in Section 3.3. In contrast, LIME assigns large importance values to Escalators 03 and 07... while most remaining escalators receive near-zero scores. Critically, the three ground-truth shut down escalators (08, 13, 14) are not consistently identified among the top-ranked features."

The paper concludes: "By incorporating in-control reference information into local surrogate construction, BREAD generates explanations that are more aligned with the variables driving out-of-control signals while preserving the model-agnosticism needed for applicability and computational efficiency needed for online deployment." Future work directions include quantifying explanation uncertainty, extending to multi-reference variants for multi-modal baseline operations, and investigating feature-selection mechanisms on top of the proposed method.

Improvements to AI Systems:

1. Anomaly Diagnosis with Baseline-Referenced Explanations

• Integrate BREAD into AI monitoring systems to replace LIME for anomaly detection. The system will generate sparse, faithful explanations that pinpoint the exact features (e.g., sensor readings, escalator IDs) driving an anomaly, even when the anomaly is far from normal behavior.

• The AI will automatically ignore noise features, assigning near-zero relevance to irrelevant variables, reducing false alarms and operator confusion.

2. Scalable Model-Agnostic Interpretation for Online Deployment

• Deploy BREAD in real-time AI pipelines (e.g., industrial IoT, network security) where traditional XAI methods fail due to computational cost. The system will provide explanations within latency constraints by using the novel sampling scheme and weighting kernel, enabling continuous monitoring without performance degradation.

3. Theoretically Guaranteed Faithfulness

• Use BREAD’s mathematical guarantee (Theorem 3.3) to build AI systems that provably concentrate relevance on true anomaly drivers as the anomaly magnitude grows. This ensures that, for large shifts, the explanation’s coefficient ratio converges to zero for irrelevant features—a property LIME lacks. The system will thus avoid misleading explanations in high-stakes scenarios (e.g., equipment shutdowns).

4. Robust and Stable Explanations Under Limited Data

• Apply BREAD in AI systems with small sample sizes (e.g., early-stage anomaly detection) to achieve high faithfulness with low variance, as demonstrated in the simulation study. The system will maintain reliable explanations even when only a few anomalous observations are available, improving trust in AI decisions.

5. Multi-Modal Baseline Handling

• Extend BREAD to multi-reference variants (as suggested in future work) to enable AI systems to diagnose anomalies across multiple normal operating modes (e.g., different production speeds or environmental conditions). The system will compare an anomaly against the most relevant baseline, yielding context-aware explanations.

6. Uncertainty-Aware Explanations

• Incorporate BREAD’s future direction of quantifying explanation uncertainty to build AI systems that output confidence intervals for feature importances. This will allow operators to know when an explanation is reliable (e.g., low variance) versus when more data is needed, improving decision-making under uncertainty.

7. Feature Selection for High-Dimensional Data

• Use BREAD’s sparsity property to implement automated feature-selection mechanisms in AI systems. The system will identify and rank only the most impactful variables for each anomaly, enabling efficient root-cause analysis and reducing manual inspection of hundreds of features.

What the Improved AI System Can Do:

• Detect and explain anomalies in real time with sparse, faithful, and stable feature attributions, even for large shifts or small datasets.

• Operate model-agnostically across black-box models (e.g., deep learning, ensembles) without retraining.

• Provide provably better explanations than LIME, especially when anomalies are severe, by ignoring noise and focusing on true drivers.

• Support online deployment in resource-constrained environments (e.g., edge devices) due to computational efficiency.

• Offer uncertainty estimates and multi-modal baseline comparisons for nuanced, trustworthy diagnostics in industrial, healthcare, or cybersecurity applications.

Artificial Intelligence (AI)-based prospective anomaly detection methods are increasingly deployed in high-dimensional and nonlinear settings. Among these approaches, AI-based statistical process monitoring (SPM) is widely used, providing a structured framework for prospective monitoring. Once an anomaly is detected, a diagnosis method is needed to identify the features driving the flagged observation away from normal behaviour. Traditional SPM diagnosis methods are typically designed for specific detection models and cannot be directly applied to AI-based methods. Model-agnostic explainable AI (XAI) offers a general framework for feature relevance explanation. However, existing methods suffer from scalability limitations or assign relevance to noise features, reducing diagnosis accuracy. We propose a scalable, baseline-referenced diagnosis method that uses both the anomalous observation and normal baseline information. We provide mathematical guarantees that under a mean-shift anomaly setting, the proposed method achieves higher faithfulness in detecting the features causing the anomaly compared to LIME. Simulation studies and a real-world case study validate the effectiveness of the proposed method and show that it generates more faithful and accurate diagnosis results for AI-based prospective anomaly detection methods.