Robust detection of hyper-local events from geotagged social media data

Authors
Xie K., Xia C., Grinberg N., Schwartz R., and Naaman M.

Architecture of our local event detection system. Including data collector, time-series builder, Gaussian Process regression model, alert engine and classifier. Arrows indicate input and output flow of each module.

Abstract
An increasing number of location-annotated content available from social media channels like Twitter, Instagram,Foursquare and others are reflecting users’ local activities and their attention like never before. In particular, we now have enough available data to start extracting real-time lo-cal information from social media. In this paper, we focus on the problem of hyper-local event detection, with the goal of enabling a monitoring and alerts system for public management officers, journalists and other users. We present a method for real-time hyper-local event detection from Instagram photos data, using two computational steps. We first use time series analysis to detect abnormal signals in a small region. We then use a classifier to decide if the detected activity corresponds to an actual event. Testing on a large-scale dataset of New York City photos, our system detects hyper-local events with high accuracy.

The research was published in Proceedings of the 13th Workshop on Multimedia Data Mining in KDD on 8/11/2013. The research is supported by the Brown Institute Magic Grant for the project CityBeat.

Access the paper: http://www.nirg.net/papers/robust-detection-sm-hyperlocal-events.pdf
To contact the authors, please send a message to mkrisch@columbia.edu