Lack of Twitter geotags can’t stop researchers from getting location

This is incredible - such an elegant insight!

For anyone too lazy to read the paper: "Our best method achieves accuracies of 64% for cities, 66% for states, 78% for time zones and 71% for regions when trained and tested using a dataset consisting of 1.52 million tweets from 9551 users from the top 100 US cities."

My first impulse was that there must be an incredible amount of noise from people tweeting passively "about" other locations (sports teams, old colleges, promoting conferences they're not attending) but they take an interesting approach in paying special to time-zones.

The problem of identifying whether someone making a tweet is traveling is an interesting factor to consider given "when users identified as travelling are eliminated, location prediction accuracy improves to 68% for cities, 70% for states, 80% for time-zones and 73% for regions."

In the paper, "a user was labeled as traveling if his/her maximum geo distance between tweets was above 100 miles, and not-traveling otherwise."

I wonder if a better approach might be to first detect if the times that they post tweets has shifted (different time-zone). If everything about a person's tweets is the same except they're categorically 2 hours later and they also tweet terms indicating a new location ("the view from x is amazing") then they could at least take a step toward pre-filtering traveling users without requiring their tweets to be geo-tagged.

For anybody else interested in geo-tagged social data, Mapbox did an incredible piece titled "Visualizing 3 Billion Tweets" here https://www.mapbox.com/blog/visualizing-3-billion-tweets/

And I submitted a great video on using geo-tagged Flickr to track bird migration patterns here: https://news.ycombinator.com/item?id=7445105

The thing that strikes me most from this article is "very few Twitter users actually tweet".