DEM: Glass-Box Framework for Interpretable Anomaly Detection in WBANs
The Distilled Explanation Model (DEM) is a new approach designed for interpretable anomaly detection in physiological sensor data obtained from Wireless Body Area Networks (WBANs). Anomalies may occur due to sensor malfunctions, network issues, or data loss, leading to false alarms. Current techniques often rely on black-box models that prioritize accuracy over transparency or utilize post-prediction explanations such as SHAP and LIME. In contrast, DEM is a three-stage glass-box framework that translates knowledge from a gradient boosting expert into a clear decision tree based on residuals from a linear baseline. This guarantees that the explanation aligns with the prediction itself. Additionally, DEM introduces a unique distillation fidelity metric to assess how well the explanation tree reflects the expert's non-linear inputs. The research is available on arXiv with ID 2605.31007.
Key facts
- DEM is a three-stage glass-box framework for interpretable anomaly detection.
- It distills knowledge from a gradient boosting expert into a decision tree.
- The decision tree operates on residuals relative to a linear baseline.
- DEM introduces a distillation fidelity metric.
- Anomalies in WBANs can be caused by sensor faults, network disruptions, or missing data.
- Existing approaches include black-box models and post-prediction methods like SHAP and LIME.
- The paper is available on arXiv with ID 2605.31007.
- The framework ensures the explanation is the prediction itself.
Entities
Institutions
- arXiv