Predicting Crime Using Twitter and Kernel Density Estimation [pdf]

ptl.sys.virginia.edu

22 points·jcr··6 comments

Interesting, but would this do any better than using real crime data, such as geolocated data police departments have already?

Even if it did worse, it could still be useful. It's common in statistical modeling that a bad model can be combined with a good model to make an even better model, and a whole ensemble of bad models can be a good model. As long as they are drawing on different data sources or are interpreting the data differently, their errors will not be exactly the same, and so they can be combined to get better results. In this case, Twitter seems like a rather different set of data than the historical crime records police departments presumably are using, and so it might be useful.

That concept is similar to the whole "wisdom of the crowd" thing, isn't it? Interesting.

Yes, wisdom-of-the-crowds or Condorcet's jury theorem are very similar.

(Now that I think about it more, there's also a lot of potential advantages to Twitter: global coverage, uniform data access, less risk of 'juking the numbers', less criteria drift over time... Yes, maybe local police records are much better, but you could spend a lifetime trying to get copies of them and putting them into the same data format and reconciling differing definitions etc.)

Wouldn't most of the information pulled just be news stations tweeting reports of crime though?

Wonder how their system would deal with #SandyLootCrew

Reply on news.ycombinator.com