A graph structure represents various networks;
such as social networks, bibliographic networks, protein-protein
interactions and communication networks.
Determining possible
links in a network is one of the most prevalent tasks which is
defined as the link prediction problem. Finding probable links
for data modeled as networks is used for predictive systems and
recommendation systems.
We compare different
supervised and unsupervised classification algorithms on their
performance to accurately predict missing links.
First, we model
our network by creating baseline parameters such as adamic
adar, resource allocation index, preferential attachment index
and common neighbors. The approach is tested on Facebook
dataset for one ego-net network. The performance of every model
is evaluated on its area under curve, receiver operating
characteristic score where we observe that Support Vector
Machine performs better than all the other supervised learning
algorithms .
Citation: J. McAuley and J. Leskovec. Learning to Discover Social Circles in Ego Networks. NIPS, 2012.