we (me and some of my colleagues) have published two years ago a paper using data-mining in the orkut (www.orkut.com) virtual community. We have chosen 2 communities (”I love beer” with 52,000 members and “I hate beer” with 32,000 members) and made a probabilistic sample of them. We have “chosen” 400 users from each of these communities.

From the users, we have collected all the data that they show at their profile, as the number of communities that they participate, how many friends, how many scraps, testimonials, and other information that he or she can describe about himself or herself.
With these, we tryed to predict if the user participate in one community or in the other

We used two different types of data analysis: 1 - logistics regression; 2 - neural networks.

we got better results with the neural network, and some of the results from this analysis are these:
Training sample: 591
validation sample: 196
variables: 121
neurons: 124 with 3 occult layers
Rsquare: 0,7446
best index in training: 95% of the cases
best index in validation: 91% of the cases

it was a quite interesting paper, unfortunatly we just have it in portuguese.