IBM Boffins KNOW WHERE YOU LIVE, thanks to Twitter
"Woohoo I'm in Sydney" tells people you're in Sydney, it seems
If you thought refraining from geotagging your Tweets or photos was enough to keep your secrets from the world at large, think again: IBM researchers say a Twitter user's primary location can be inferred from their behaviour, with accuracy as high as 68 per cent.
In this paper at Arxiv, Jalal Mahmud, Jeffrey Nichols and Clemens Drews of IBM Research at Almaden say they can at least get city-level predictions of Twitter users' “home” locations (by which they mean the primary location from which an individual usually Tweets), even though the user isn't using Twitter's location features.
To do this, the researchers produced two algorithms. The first uses behaviours such as volume of Tweets from a user, and external information (a dictionary of location names and services such as Foursquare). They say that while this algorithm works best when users make “explicit references” of locations in Tweets, it “still works with reduced accuracy when no explicit references are available”.
The second algorithm predicts locations “hierarchically using time zone, state or geographic region as the first level and city at the second level”.
With a dataset of around 1.5 million Tweets from 9,551 users, the researchers then extracted classifiers including:
- All words in the Tweets;
- All hashtags in the Tweets; and
- All city and state location names in the Tweets.
Armed with this data, the researchers then note, they can also make some assumptions about location – for example, given America's timezones, a user in New York is more likely to be at home at 7:00PM eastern time, while at the same time, a Californian user is probably still at work. That means a user's volume of Tweets helps become a hint to their location.
The paper notes that “geo-tags are not used in any of our prediction algorithms, although around 65 per cent of the tweets in our dataset are geo-tagged”.
But don't worry, the researchers only intend their work to be used for good: “a journalist tracking an event on Twitter may want to know which tweets are coming from users who are likely to be in a location of that event, vs. tweets coming from users who are likely to be far away. As another example, a retailer or a consumer products vendor may track trending opinions about their products and services and analyse differences across geographies.
“Second, our examination of the discriminative features used by our algorithms suggests strategies for users to employ if they wish to micro-blog publicly but not inadvertently reveal their location”, the study notes. ®