“Each [Kaggle] contest has generated a better model for its data than what was used beforehand.” (Science)
“No snake-oil. no bullshit.”
Following up on my previous post on scientific crowdsourcing it is really interesting to see where they are taking this – I myself got the “bug” when participating in the IJCNN social network contest. By using relatively simple network-internal descriptors (i.e. common neighbours in the 1st and 2nd shell) I landed somewhere in the mid-field. The top-scorers used de-anonymizing the network and crawling flickr – such options didn’t even cross my mind. So I learned something from the winning strategies by reading up on the papers and blog entries on the methods they employed – as described in this talk. On the one hand it demonstrates the power of mapping networks for computational inference, on the other hand it also taught me to sometimes think “outside” the given data. Needless to say, I like to look at things from a different angle – so its nice to see it working as a way of scientific discovery for others as well.
see also this special on ABC catalyst.