Motivation

Our Contributions
Summary
Better Approximation of Betweenness
Centrality
Robert Geisberger Peter Sanders Dominik Schultes
Workshop on Algorithm Engineering & Experiments, 2008
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Motivation
Automatic analysis of networks requires fast computation of
centrality indices.
The networks grow faster than the speed of our computers so
fast approximation algorithms gain importance.
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Transportation
applications:
e.g. routing
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Graph drawing
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Denition Betweenness Centrality
Let
G = (V, E) be a weighted directed (multi)-graph,
SP
st
= set of shortest paths between source s and target t
SP
st
(v) = set of shortest paths that have v in their interior.
Then the betweenness centrality for node v is
c(v) :=

s,t V

st
(v)

st
, where
st
:= |SP
st
| and
st
(v) := |SP
st
(v)| .
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Exact algorithm
Brandes [Brandes01] exact algorithm:
solve single source shortest path problem (SSSP) from
each node
backward aggregation of counter values
s v
w
w

s v
w
w

+ =
s v
w
w

Time requirements:
(nm) for unit distance, otherwise

_
nm +n
2
log(n)
_
.
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Approximation approach
Brandes and Pich [BrandesPich06] approximation algorithm:
choose subset k of starting nodes (pivots)
solve only k single source shortest path problem (SSSP)
extrapolate betweenness values
This yields an unbiased estimator for betweenness.
pivot
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Computation of centrality indices
Previous Work
Deciency of previous approach
Overestimation of betweenness values of nodes near a pivot.
pivot
large overestimation small overestimation "false zero"
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Main idea
Consider the length to the pivot to scale contributions.
pivot
large overestimation small overestimation "false zero"
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Generalized Framework
Parameters:
length function on the edges
For a path P = e
1
, . . . , e
k
let (P):=

1i k
(e
i
)
scaling function f : [0, 1] [0, 1]
Features:
unbiased estimator
focus on differences between approximation methods
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Generalized Framework (continuation)
For each shortest path of the form

Q
..
s, . . . , v, . . . , t
. .
P

we dene a scaled contribution

P
(v):=
f ((Q)/(P))

st
Overall, v gets a contribution from a pivot s

s
(v):=

t V

{
P
(v) : P SP
st
(v)}
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Proposed Parameters
Length function :
edge weight function used for shortest-path calculation
unit distance
Scaling function f :
Brandes and Pich constant
linear scaling f (x) = x
bisection scaling
f (x) =
_
0 for x [0, 1/2)
1 for x [1/2, 1]
0 1
0
1
0 1 0 1
Brandes and Pich linear scaling bisection scaling
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Linear Time Computation
Brandes [Brandes01]:
compute
st
on the y during the shortest path calculation
subsequent aggregation phase, like exact algorithm
linear scaling:
Let
st
denote the shortest path distance from s to t ,
aggregate 1/
st
instead of 1, multiply with
sv
at the end.
s
t
t

v
w
1

st
1

st

sv

sw
+
1

st
+
1

st

sw

st
+
1

st

Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Linear Time Computation of bisection scaling
use unit distance
depth rst traversal of shortest path DAG, keep an array
storing the current path from s
increment counter of current node v and decrement
counter of middle node v

s
v

v
0

d
2

1 d

f
+1 1
Comments:
only efcient for
st
{0, 1}
for
st
2 sampling of shortest paths required
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Overview of used graphs
graph nodes edges source
Belgian road network 463 514 596 119 PTV AG
Belgian road network (unit dist.) 463 514 596 119 PTV AG
Actor co-starring network 392 400 16 557451 [NotreD]
US patent network 3 774769 16 518947 [NBER]
World-Wide-Web graph 325 729 1 497135 [NotreD]
CNR 2000 Webgraph 325 557 3 216152 [LabWA]
CiteSeer undir. citation network 268 495 2 313294 [Citeseer]
CiteSeer co-authorship network 227 320 1 628268 [Citeseer]
CiteSeer co-paper network 434 102 32 073440 [Citeseer]
DBLP co-authorship network 299 067 1 955352 [DBLP]
DBLP co-paper network 540 486 30 491458 [DBLP]
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Belgium road network
0
.
5
1
2
3
0
.
5
1
2
3
16 32 64 128 256 512 1024 2048 4096 8192
pivots for Brandes' algorithm
%

o
f

p
o
s
s
i
b
l
e

i
n
v
e
r
s
i
o
n
s
Brandes
bisection (unit)
bisection (sh.path)
linear
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Generalized Framework
Efcient Implementation
Experiments
Belgium road network
0
.
5
0
.
2
1
2
0
.
5
0
.
2
1
2
0 1 2 3 4 5 6 7 8 9 10 11
level
r
e
l
a
t
i
v
e

r
a
n
k

e
r
r
o
r
Brandes
bisection (unit)
bisection (sh.path)
linear
0.1%
1%
10%
25%
50%
75%
90%
99%
99.9%
least important vertices most important vertices
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Motivation
Our Contributions
Summary
Summary
The bisection scaling algorithm achieves the best results.
Future work
efcient exact bisection scaling algorithm for
st
2
local searches to eliminate "false zeros"
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Appendix Additional Experiments
Belgian road network
1
0

4
1
0

3
1
0

2
1
0

4
1
0

3
1
0

2
16 32 64 128 256 512 1024 2048 4096 8192
pivots for Brandes' algorithm
E
u
c
l
i
d
e
a
n

d
i
s
t
a
n
c
e
Brandes
bisection (unit)
bisection (sh.path)
linear
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Appendix Additional Experiments
Belgian road network (Brandes and Pich)
2
3
2
3
2
4
2
5
2
6
2
7
2
8
2
9
2
10
2
11
2
12
2
13
2
14
2
15
2
16
2
17
2
18
pivots for Brandes' algorithm
%

o
f

p
o
s
s
i
b
l
e

i
n
v
e
r
s
i
o
n
s
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality
Appendix Additional Experiments
Additional networks
1
5
2
1
0
2
0
1
5
2
1
0
2
0
16 32 64 128 256 512 1024 2048 4096 8192
pivots for Brandes' algorithm
c
o
m
p
a
r
e

i
n
v
e
r
s
i
o
n

B
r
a
n
d
e
s

/

b
.
s
a
m
p
l
.
(
2
)
Belgium unit
Belgium
DBLP copaper
CiteSeer citation
CiteSeer copaper
DBLP coauthor
CiteSeer coauthor
Robert Geisberger, Peter Sanders, Dominik Schultes Better Approximation of Betweenness Centrality