class 12: communities network science: communities dr. baruch barzel
TRANSCRIPT
Class 12: Communities
Network Science: Communities
Dr. Baruch Barzel
The Modular Structure of Networks
)(kP
k
C
ijD
Is a Network Modular
Clustering implies modularity
Small Worldness tends to wipe out modularity
Functionality requires modularity
Is a Network Modular
Clustering implies modularity
Small Worldness tends to wipe out modularity
Functionality requires modularity
Hubs tends to wipe out modularity
Is a Network Modular
Clustering at the periphery only Low degree nodes typically belong to a single module
Hubs bridge between different modules
C
k
Is a Network Modular
Clustering at the periphery only Low degree nodes typically belong to a single module
Hubs bridge between different modules
C
k
But how do we unveil the modules
The Modular Structure of Networks
Functional modularityNatural partition lines
Network Partitioning
Optimally dividing the network into a predefined number of partitions
nx
x
x
2
1
Dividing a task into sub-tasks
Network Partitioning
Optimally dividing the network into a predefined number of partitions
nx
x
x
2
1
Dividing a task into sub-tasks, while minimizing the transmission between tasks
Network Partitioning
Optimally dividing the network into a predefined number of partitions
nx
x
x
2
1
Dividing a task into sub-tasks, while minimizing the transmission between tasks
Network Partitioning
Minimizing the Cut:
)(2
1
jQiijAR
The index vector:
bluein is if 1
redin is if 1
i
isi
ij
ijji AssR 12
1
2
1
sLssAksRij
Tjijijii
/ 4
1
4
1
ij i
ijijiiij ksskA The Laplacian Matrix:
Otherwise
0 and
0
1
ij
i
ij Aji
jik
L
The Laplacian Matrix
Minimizing the Cut: sLsR T /4
1
Otherwise
0 and
0
1
ij
i
ij Aji
jik
L
Consider the Eigenvector: ssL/
ssR T 4
1
Choose the Eigenvector with the minimal Eigenvalue
The Laplacian Matrix
Minimizing the Cut: sLsR T /4
1
Consider the Eigenvector: ssL/
ssR T 4
1
Choose the Eigenvector with the minimal Eigenvalue
negative iselement if
positive iselement if
1
1
is
The Laplacian Matrix
The matrix: ijijiij AkL
0)( iij
ijijij
jij kkAksLsL /
The trivial partitioning – put the entire network together:
ors
The Laplacian Matrix
The matrix: ijijiij AkL
The case of isolated components
The number of Eigenvectors with λ = 0 equals the number of connected components
L/
s
The Laplacian Matrix
The matrix: ijijiij AkL
The case of almost isolated components
The Eigenvectors with λ close to zero capture the partitioning
L/
s
From Partitioning to Communities
The number of communities and their size should be given by the network itself.
Hierarchical Clustering
4123
1410
2143
3034
ijW
Edges 4 Sides Stable Equal1. Square + + + +
2. Rectangle + + + --3. Circle -- -- -- --4. Triangle + -- + +
4
Hierarchical Clustering
4123
1410
2143
3034
ijW
Edges 4 Sides Stable Equal1. Square + + + +
2. Rectangle + + + --3. Circle -- -- -- --4. Triangle + -- + +
4
3
Hierarchical Clustering
4123
1410
2143
3034
ijW
Edges 4 Sides Stable Equal1. Square + + + +
2. Rectangle + + + --3. Circle -- -- -- --4. Triangle + -- + +
4
3
2
Hierarchical Clustering
4123
1410
2143
3034
ijW
Edges 4 Sides Stable Equal1. Square + + + +
2. Rectangle + + + --3. Circle -- -- -- --4. Triangle + -- + +
4
3
2
1
Dendograms
Dendograms
Topologically Induced Weights
jiWij and between paths ofnumber the toRelated
1
0
)()(
AIAWl
l
pathst independen node ofNumber ijW
jiA ijl and between paths ofNumber
Betweeness
Edge Betweeness – the number of paths through an edge
Football and Karate Networks
Zachary’s Karate Club
College Football
Football and Karate Networks
Zachary’s Karate Club
College Football
Ising and Potts Models
ji
ssij jiJE
,
0T T
1is
Ising and Potts Models
ji
ssij jiJE
,
0T T
Groups of nodes with high link density will tend to have the same polarization
Sparseness of connections between groups will allow different communities to have unrelated spins
1is
Ising and Potts Models
ji
ssij jiJE
,
0T T
Groups of nodes with high link density will tend to have the same polarization
Sparseness of connections between groups will allow different communities to have unrelated spins
mmmsi ,,1, Potts
Model
Ising and Potts Models
ji
ssij jiJE
,
0T T
0T
T
mmmsi ,,1,
Ising and Potts Models
ji
ssij jiJE
,
0T T
0T
T
mmmsi ,,1,
Ising and Potts Models
ji
ssij jiJE
,
0T T
0T
T
mmmsi ,,1,
Ising and Potts Models
ji
ssij jiJE
,
0T T
0T
T
mmmsi ,,1,
Ising and Potts Models
ji
ssij jiJE
,
0T T
0T
T
mmmsi ,,1,
Link Communities
Community - A group of densely connected nodes
A group of topologically similar links
Project Presentations (5 min.)
1. Define your network (nodes, links)
2. How will you get the data
3. Estimated size of network
4. Why