We Need to Rethink Benchmarking in Anomaly Detection
A new position paper argues that current benchmarking methods for anomaly detection algorithms are inadequate and hinder progress. The paper highlights that simple algorithms often perform competitively with advanced deep learning models, and existing benchmarks do not reflect the diverse real-world applications of anomaly detection. To address this, the authors propose a new approach that groups applications into scenarios based on a common taxonomy, allowing for scenario-specific choices in preprocessing, metrics, and model selection to provide better guidance for practitioners. AI
IMPACT Proposes a new framework for evaluating anomaly detection models, potentially improving practitioner guidance and driving more meaningful algorithmic advancements.