Enhanced Detection of Crisis-Related Microblogs by Spatiotemporal Feedback Loops

Microblogs such as Twitter tweets are an important information source in emergency response and management. However, current approaches are either not accurate enough in identifying relevant crisis events (e.g., explosion, tornado) or are limited to detect a single event type. In this paper, we present a spatiotemporal feedback loop extension for boosting existing (black-box) detection approaches by re-labeling misclassified microblogs in geographical topic hotspots afterward. To show the applicability, we evaluate our approach on 6 publicly available (geo-tagged) datasets and compare it with the standard (black-box) approach without spatial extension. Our results show significant F1-score performance gains up to 5.8% only by enabling our spatiotemporal extension.