Local, multi-resolution detection of network communities by Markovian dynamics
File(s)
Author(s)
Yu, Yun William
Type
Thesis
Abstract
Complex networks are used to represent systems from many disciplines,
including biology, physics, medicine, engineering and the social sciences;
Many real-world networks are organised into densely connected communi-
ties, whose composition gives some insight into the underlying network.
Most approaches for nding such communities do so by partitioning the
network into disjoint subsets, at the cost of requiring global information
and that nodes belong to exactly one community. In recent years, some effort
has been devoted towards the development of local methods, but these
are either limited in resolution or ignore relevant network features such as
directedness.
Here we show that introducing a dynamic process onto the network allows
us to de ne a community quality function severability which is inherently
multi-resolution, takes into account edge-weight and direction, can accommodate
overlapping communities and orphan nodes and crucially does not
require global knowledge. Both constructive and real-world examples|
drawn from elds as diverse as image segmentation, metabolic networks
and word association|are used to illustrate the characteristics of this approach.
We envision this approach as a starting point for the future analysis
of both evolving networks and networks too large to be readily analysed as
a whole (e.g. the World Wide Web).
including biology, physics, medicine, engineering and the social sciences;
Many real-world networks are organised into densely connected communi-
ties, whose composition gives some insight into the underlying network.
Most approaches for nding such communities do so by partitioning the
network into disjoint subsets, at the cost of requiring global information
and that nodes belong to exactly one community. In recent years, some effort
has been devoted towards the development of local methods, but these
are either limited in resolution or ignore relevant network features such as
directedness.
Here we show that introducing a dynamic process onto the network allows
us to de ne a community quality function severability which is inherently
multi-resolution, takes into account edge-weight and direction, can accommodate
overlapping communities and orphan nodes and crucially does not
require global knowledge. Both constructive and real-world examples|
drawn from elds as diverse as image segmentation, metabolic networks
and word association|are used to illustrate the characteristics of this approach.
We envision this approach as a starting point for the future analysis
of both evolving networks and networks too large to be readily analysed as
a whole (e.g. the World Wide Web).
Version
Open Access
Date Issued
2013-08
Date Awarded
2014-04
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Barahona, Mauricio
Yaliraki, Sophia
Sponsor
Marshall Scholarships
Publisher Department
Mathematics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Master of Philosophy (MPhil)