We address the problem of multi-label classification of relational
graphs by proposing a framework that models the input graph as a first
order Markov random field and devises a relaxation labeling procedure to
find its maximally likely labeling. We apply this framework to
classification as well as clustering problems in homogeneous networks
and show significant performance gains in comparison to state-of-the art
techniques.
We also address the problem of multi-label classification in
heterogeneous networks where every data point is associated with a node
type and has to be labeled with one or more classes from a type-specific
finite set of classes. Our algorithm is based on a random walk model. We
present detailed empirical studies of our model and compare it with
state-of-art techniques on two social networks.
All newly proposed algorithms are robust to scarce training data and
diverse linkage patterns.
They improve classification or clustering quality in homogeneous and
heterogeneous networks.