| Popis: |
Accurately detecting instances in datasets that have been mislabeled is a difficult problem with several imperfect solutions. Hand-reviewing labels is a reliable but expensive approach. Time series datasets present additional challenges because they are not as easily interpreted by reviewers. This paper introduces TSAR, as system for facilitating human review of a small portion of a dataset that it identifies as the most likely to be mislabeled. TSAR’s use is demonstrated on real-world time series data. |
