Range-based metrics

[LouisCarpentier42]

Implement Range-based precision, recall and F1 [1] to evaluate anomaly detection performance. Traditional precision and recall consider each point independently, and therefore do not take temporal dependencies in time series into account. People have proposed point-adjusting the predicitons, in which an entire anomalous event is considered detected, if any of the individial points are predicted to be an anomaly. This point-adjustement does, however, largely overestimate the performance of anomaly detectors and may lead to random detectors outperforming state-of-the-art. The range-based metrics proposed in [1] overcome such issues by considering the temporal nature of time series.

[1] N. Tatbul, T. J. Lee, S. Zdonik, M. Alam, and J. Gottschlich. Precision and recall for time series. Advances in neural information processing systems, 31, 2018.