Shared publicly  - 
 
Social Relationship Tracking: You Can Run, But You Can’t Hide
Machine learning team demonstrates how to infer data about non-members of a social network
by +Sophie Wrobel, +EuroTech; Germany

Are you concerned about what the social networks that you are not a member of know about you? Maybe you should be. If Facebook gets their algorithms right, they could make a surprisingly good guess about whether two non-Facebook users are friends with each other. Should you be concerned?

According to a team of researchers from the University of Heidelberg, social networks can predict information about the relationship between a member and a non-members of a given social network platform with 85% accuracy, and information about the relationship between two non-members of the same network with 40% accuracy. Their prediction model works by testing various parameterized neural networks across a set of realistic Facebook training data in order to answer the question of whether a particular non-user is likely to be recruited by a particular user.

But perhaps more important than their methodology are the repercussions of their results. Even when commercial applications of their research are still way off, this will likely turn into another case of balancing out the marketer’s dream with the individual user’s nightmare: how much inferred data should be allowed in targeting advertisements? What non-personal hints can and should be allowed to be used in calculating marketing target probabilities?

How anonymous can an incognito user truly be in light of such elaborate algorithms?

More information:
http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0034740

___________________________________________________
Show your support by plussing the page and sharing it with friends.
→ More +EuroTech news | http://zipl.us/eurotech/feed (RSS)
14
11
Alexander Becker's profile photoFrançois-Xavier Thomas's profile photo
3 comments
 
Just as in real life, you can hardly guard against intelligent inference. For those who are paranoid enough though, 40% probability, while great as a scientific result, is nothing to write home about in tangible terms.
 
+Alexander Becker : 40% probability is even less than what you'd get by flipping a coin ;)
But still, mathematically at least, it's a very interesting theory!
 
+David Oliver : From where I stand, usually, when you've got a predictive theory, you can only start applying it to concrete problems when you get to about 80%-85% recognition rate.

The interesting thing is, it all depends on what you need. In image processing, we have this thing called "precision-recall curves", where, basically, the precision is the amount of true positives in your results, and the recall is the amount of real things you've retrieved.

Sometimes, you want to have more precision : if you're doing OCR on zip codes written on postcards, for instance, you want to immediately send the postcards where are sure you recognize the zip code, and send the remaining postcards for classic human-based processing. You can do so by increasing precision, but then you get fewer cards that are sent without the need for human help.

On social networks, on the other hand, the precision component is much less relevant, so you can allow for much less precise algorithms and still be happy about it. So, you can decrease the precision to improve the recall, the amount of predicted connected users.

So, you can have an algorithm that is very inefficient or imprecise, and people can still use it because, in their specific framework, it works ;)

(If you want a visual example from a real case, take this : http://www.cs.cmu.edu/~hebert/RS/evaluateAllPrecisionRecall_outlier2.png : the higher the curve is, the better the algorithm works)
Add a comment...