Anomaly Explanation Using Metadata

摘要: Anomaly detection is the well-studied task of identifying when data atypical in some way with respect to its source. In this work, by contrast, we are interested finding possible descriptions what may be causing anomalies. We propose a new task, attaching semantics drawn from metadata portion anomalous examples Such partial description terms meta-data useful both because it help explain causes identified anomalies, and also identify truly unusual that defy such simple categorization. This especially significant set too large for human analyst inspect anomalies manually. The challenge are, definition, relatively rare, so seeking learn precise characterization rare event. examine algorithms webcam domain, generating human-understandable explanations pixellevel find using recently proposed algorithm prioritizes precision over recall, attach good moderate fraction long as fairly large.