Metro-Viz
Philipp Eichmann, Franco Solleza, Junjay Tan, Nesime Tatbul, Stan Zdonik · 2019
Millions of time-based data streams (a.k.a., time series) are being recorded every day in a wide-range of industrial and scientific domains, from healthcare and finance to autonomous driving. Detecting anomalous behavior in such streams has become a common analysis task for which data scientists employ complex machine learning models. Analyzing the behavior and performance of these models is a challenge on its own. While traditional accuracy metrics (e.g., precision/recall) are often used in practice to measure and compare the performance of different anomaly detectors, such statistics alone are insufficient to characterize and compare the algorithms in a systematic, human-interpretable way. In this extended abstract, we present Metro-Viz, a visual analysis tool to help data scientists and domain experts reason about commonalities and differences among anomaly detectors, and to identify their strengths and weaknesses.