New Framework Categorizes Time Series Anomaly Detection Metrics
A study introduces a problem-oriented taxonomy for evaluating time series anomaly detection, reinterpreting over twenty metrics across six dimensions: accuracy, timeliness, labeling tolerance, audit cost, robustness, and comparability. The framework moves beyond mathematical forms to address specific evaluation challenges in IoT and cyber-physical systems. Experiments compare metric behavior under genuine, random, and oracle scenarios.
Key facts
- The study introduces a problem-oriented framework for time series anomaly detection evaluation.
- Over twenty commonly used metrics are categorized into six dimensions.
- The six dimensions are: basic accuracy, timeliness, labeling imprecision tolerance, human-audit cost penalties, robustness against random scores, and parameter-free comparability.
- Comprehensive experiments examine metric behavior under genuine, random, and oracle detection scenarios.
- The framework reinterprets metrics based on evaluation challenges rather than mathematical forms.
- The research is published on arXiv with ID 2511.18739.
- Time series anomaly detection is widely used in IoT and cyber-physical systems.
- The study aims to address challenges due to diverse application objectives and heterogeneous metric assumptions.
Entities
Institutions
- arXiv