Active Learning in Link Prediction for social networks
This is just a draft version for the project idea and it is in a partner hunting state.
Contents
Members
And I am looking for partners.
Idea
In most online social networks, link prediction is often used as a friend suggestion that users make confirmation to a list of friend suggestions from the system. The feedback of users can be very informative as it sometimes tells the bias of this user, especially temporal bias or interest. And this scenario is similar to an active learning framework that the system can query the user by providing a suggestion and learn from the feedback. We can also see this feedback in a learning-to-rank model that users' confirmation on the suggestions as a click event on the search results. Active learning in link prediction intuitively also should work well in a cold start situation due to lack of information and active learning can quickly converge and learn more about the user from informative suggestions feedback.
Data sets
Stanford data sets and a lot of social network data from datamob, which has a 2009 facebook social graph data. And I also have a protein-protein interaction network dataset. Although it is not social network, it shares some important facts like small-world phenomenon and a power low distribution of node degree and modularity. It can be used to test on a cold start situation as the average degree is only 4 to 5.
Evaluation Metrics
Evaluation is an interesting topic in social network link prediction. One common way is to divide the graph by different timestamps and train on an early graph and test the prediction on a later one. This sounds natural, however, sometimes and especially in a new social network, the sequence of adding friends is just based on the how the system gives the suggestion rather than some real features we can capture. And we should have a more suitable way to deal with our active learning framework.
