[Paper Review] Inferring the geographic focus of online documents from social media sharing patterns

Ryan Compton, Matthew S. Keegan+1 more|arXiv (Cornell University)|Jun 10, 2014

Geographic Information Systems Studies33 references3 citations

TL;DR

This paper proposes a content-agnostic method to infer the geographic focus of online documents by analyzing the geolocations of Twitter users who share links to them. Using robust statistical estimation (median and median absolute deviation) on shared-user locations, the method achieves high accuracy for tweets and YouTube videos, with mean error controllable via dispersion thresholds, enabling scalable, low-resource geotagging without content analysis.

ABSTRACT

Determining the geographic focus of digital media is an essential first step for modern geographic information retrieval. However, publicly-visible location annotations are remarkably sparse in online data. In this work, we demonstrate a method which infers the geographic focus of an online document by examining the locations of Twitter users who share links to the document. We apply our geotagging technique to multiple datasets built from different content: manually-annotated news articles, GDELT, YouTube, Flickr, Twitter, and Tumblr.

Motivation & Objective

• To address the scarcity of publicly available geographic annotations in online digital media.
• To develop a content-agnostic geotagging method that does not require natural language processing or image analysis.
• To evaluate whether social media sharing patterns can reliably infer the geographic focus of online content.
• To quantify the accuracy and uncertainty of geotagging using only link-sharing user locations.
• To demonstrate the method's effectiveness across diverse content types: news, tweets, YouTube, Flickr, and Tumblr.

Proposed method

• The method computes the geographic geotag of a URL as the L1-multivariate median of the geolocations of Twitter users who shared the URL.
• It uses Vincenty’s formulae to compute geodesic distances between locations for median computation.
• A dispersion threshold based on median absolute deviation (MAD) is applied to filter out low-confidence geotags.
• The method is applied to URLs from news, YouTube, Flickr, Twitter, and Tumblr, using Twitter user location data as input.
• Robust statistics are essential to mitigate the impact of geographic outliers in social media data.
• The approach is evaluated by comparing inferred geotags to ground-truth locations from official APIs (YouTube, Flickr) or manual annotations (news, toponyms).

Experimental results

Research questions

• RQ1 Can the geographic focus of online documents be accurately inferred from the locations of Twitter users who share links to them?
• RQ2 How does the performance of content-agnostic geotagging vary across different content types such as news, tweets, YouTube videos, and photos?
• RQ3 To what extent does dispersion filtering improve geotagging reliability and accuracy?
• RQ4 How does the method compare to content-specific geotagging techniques in terms of coverage and error?
• RQ5 Why does the method perform poorly on Tumblr reblogs despite similar sharing patterns?

Key findings

• The method achieves high accuracy in inferring geographic focus for tweets, with performance nearly an order of magnitude better than for news articles.
• For YouTube videos, the method inferred geotags for 447,441 URLs, while only 14,582 (3%) were annotated via the YouTube API, demonstrating high scalability.
• On Flickr, only 284 out of 1,992 URLs had official geotags, indicating widespread lack of public location metadata.
• The mean error in YouTube geotagging can be reduced by half while retaining 60% coverage by applying a dispersion threshold of 1000 km.
• The method performs poorly on Tumblr, where reblogs are not geographically grounded, as shown by high discrepancy between inferred and toponym-based geotags.
• Dispersion filtering via median absolute deviation effectively reduces uncertainty and improves reliability, especially for high-variability content like news and general web content.