Rafael Lima
Samsung R&D Institute Brazil
Avenida Cambacicas 1200
Campinas-SP, Brazil
[email protected]
Chris Solomou
University of York
Deramore Lane Heslington York
York, UK
[email protected]
Temporal Point Processes have undergone increasing relevance in the modeling of continuous-time event streams. Regarding their applicability, one important aspect is that of detecting anomalous, or out-of-distribution, sequences. Recent works have focused on parametric models for this out-of-distribution detection. In the present work, we give a theoretical background treatment of the anomaly detection problem applied to TPPs, describe our fully neural-based strategy, show how a fully neural-based strategy of improved generalization outperforms traditional parametric approaches, and validate its effectiveness against a state-of-the-art approach on data of controlled generation. https://openreview.net/pdf?id=7opQsdC2Ob
