Communities, Spectral Clustering, and Random Walks
David Bindel (CS, Cornell)
We think of a community in a social network as a subgraph that is
unusually tightly connected. But there are many possible definitions
of what it means to be unusually tightly connected
, and these lead
to a plethora of formal definitions of communities in terms of
subgraphs that minimize or maximize some metric. Several of these
metrics can be expressed in terms of quadratic forms, and so relaxed
versions of these optimization problems lead naturally to eigenvalue
problems and to spectral methods to community detection. After
discussing some standard spectral ideas that produce detect disjoint
communities, we discuss a new approach to finding overlapping
communities by a constrained l1 minimization problem over a subspace.
We then turn to another approach to community detection based on the
intermediate dynamics of random walks. Though superficially
dissimilar to the spectral approach, we show that this
random-walk-based method is fundamentally related to standard iterations for
computing eigenvectors. We also comment on a curious connection between
community detection via random walks and the dynamics of transitions between
quasi-equilibrium states in chemical physics.