Multiscale event detection using convolutional quadtrees and adaptive geogrids

Abstract

Increasing popularity of social networks made them a viable data source for many data mining applications and event detection is no exception. Researchers aim not only to find events that happen in networks but more importantly to identify and locate events occurring in the real world. In this paper, we propose an enhanced version of quadtree - convolutional quadtree (ConvTree) - and demonstrate its advantage compared to the standard quadtree. We also introduce the algorithm for searching events of different scales using geospatial data obtained from social networks. The algorithm is based on statistical analysis of historical data, generation of ConvTrees representing the normal state of the city and anomalies evaluation for events detection. Experimental study conducted on the dataset of 60 million geotagged Instagram posts in the New York City area demonstrates that the proposed approach is able to find a wide range of events from very local (indie band concert or wedding party) to city (baseball game or holiday march) and even country scale (political protest or Christmas) events. This opens up a perspective of building a simple and fast yet powerful system for real-time multiscale events monitoring.