response network emerging from simple perturbation
DESCRIPTION
Response network emerging from simple perturbation. Seung-Woo Son Complex System and Statistical Physics Lab., Dept. Physics, KAIST, Daejeon 305-701, Korea. Motivation : Microarray Data. - PowerPoint PPT PresentationTRANSCRIPT
Response network emerging from simple perturbation
Seung-Woo Son
Complex System and Statistical Physics Lab.,
Dept. Physics, KAIST, Daejeon 305-701, Korea
Motivation : Microarray Data
• Microarray data show the response of each gene to an experiment, which is a kind of perturbation to the genetic network. ex) gene deletion, temperature change etc
• Like building the genetic network from microarray data, the secondary network can be constructed from the response of primary network under perturbation. ex) node removal (?)
“ Can the secondary network represents the primary network correctly ? ”
“ What is the meaning of the response under perturbation ? ”
“ Ultimately, can we find out primary network from the secondary network ? ”
Introduction : Node Removal Perturbations
• When a node is removed, network structure changes. The network can break into several isolated clusters.
• Giant cluster size decreases gradually and the average path length increases.
R. Albert and A.-L.Barabási, Reviews of Modern Physics, 74, 47 (2002)
• SF network is more tolerant against random removal better than random network.
• In SF network, the diameter changes under a node removal follow the power-law distributtion. J.-H. Kim, K.-I. Goh, B. Kahng and D. Kim,Physical Review Letters, 91, 5 (2003)
Introduction : Load & Betweenness Centrality
• What is the “Load” ?– When every pair of nodes in a network exchanges data
packets along the shortest path, load of a node is the total number of data packets passing through that node.
ex) Internet traffic jam
ji j
k11
2
1
2
1
4
1
4
3
ji
ji klkl,
)()(
start target
i
k11
3
2
3
1
3
1
3
2
ji
ji kbkb,
)()(
start target
• Betweenness Centrality BC ( Freeman, 1977 )
– if is the number of geodesic paths from i to j and is the number of paths from i to j that pass through k, then is the proportion of geodesic paths from i to j that pass through k. The sum for all i,j pairs is betweenness centrality.
jig k
jig
jik
ji gg /
jik
jiji ggkb /)(
)(kl ji
1000 10000
1
10
100
1000 bi (avg. over row)
b(k) (avg. over column)
~ 2.0
Dis
trib
utio
n of
b
i ,
b(k)
bi & b(k)
Introduction : BC Changes . - BA model
)(kbi
)()()( kbkbkb oii
network wholeafter the nodeth -k of BC
removal nodeth -iafter nodeth -k of BC )(
o
i
b
kb
)()1(
)()1( 11
jbb
jbb
bii
cf) diameter changes
J.-H. Kim, K.-I. Goh, B. Kahng and D. Kim,Physical Review Letters, 91, 5 (2003)
Distribution of . - BA model)(kbi
o
ii b
ibb )(
• distribution is power law distribution with exponent 2.1
)(kbi
• Summation of BC changes after i-th node removal is linearly proportional to BC of i-th node in BA model.
)(ibb oi )()()()( ibbbkbkbkbb o
k k koioiii
MST & Percolation Network• How to build the secondary network ? : Based on = “correlation” bewtween node i and j
– MST (minimum spanning tree)A graph G = (V,E) with weighted edges. The subset of E of G of minimum weight which forms a tree on V ≡ MST .A node is linked to the most influential one with constraint such that N vertices must be connected only with (N-1) edges.
– PercolationAfter sorting Δbi(j) in descending order, add a link between i and j following that order. When all nodes make a giant cluster, stop the attachment. It means the links with values Δbi(j) > b* (percolation threshold) are valid and connected.
)()1(
)(. 1
jbb
jb
bii
a
cb
de
f
3 4
5
4
2
6
31
2
7
MST
)( jbi
a
cb
de
f
3 4
5
4
2
6
31
2
7
Percolation
Result : Secondary Networks
• The degree k of secondary networks contain the global information of primary network, because it is constructed from BC that is calculated from the information of whole network.
• More sparse or dense networks which contain the information of original network can be constructed.
• Secondary networks represent the primary network well with significant link matches.
BA 100 MST
Secondary network construction
Result : Minimal Spanning Tree
1 10 100
1
10
100
1000BA model ( N = 1000, m = 2 )
~ 2.2
kMST
exponet 2.2 fit line
dist
arib
utio
n D
( k m
st )
degree of MST ( kmst
)
-10 0 10 20 30 40 50 60 70 80-20
0
20
40
60
80
100
120
140
160
k mst ,
k pe
r
degree of primary network ( korg
)
mst percolation original slope 1 slope 3.3 slope 0.9
• In MST network, the degree distribution shows the power-law with exponent 2.2 not 3.0 ( Scale-free )
• The degree of each node in secondary network is linearly correlated to that of primary network.
orgpermst kkk ,
2.2~)( mstmst kkD
Result : Percolation Network
• The degree distribution of percolation network shows power-law. ( exponent -1.9 )
• Percolation features appear during giant cluster fromation.
1 10 100
1
10
100
BA model ( N = 1000, m = 2 )
percolation k exponent 1.9 fit line
~ 1.9
dist
ribut
ion
D (
kp
)
degree of percolation network ( kp )
1E-3 0.01 0.1 1
0.0
0.2
0.4
0.6
0.8
1.0
size
of
gian
t cl
uste
r (
G /
N )
percolation value ( v / vmax
)
BA 1000 nodes BA 3000 nodes
1 10 100 10001
10
100
1000
~ 0.9
number of nodes (n) size of giant cluster (s) linear fit 0.91
nu
mb
er
of
no
de
& s
ize
of
gia
nt
clu
ste
r
number of link attached (m)
Similarity Measurement between Two Networks
• The links of each node are regarded as vector in N dimensional vector space.– Vector inner product shows the
similarity between two networks.
• Binary undirected network case : It means how many links are overlapping each other.
1
2
3 0
45
6
1
2
3 0
45
6 compare
),,,,,( 21 iNijiii wwwwv 'iii vvx
N
1i
X measure similarity ix
BA model ( N = 1000 , M = 1996 )
links Xmatche
s
MST network 999 0.908 907
Percolation net.
3377 0.766 1529
Other BA net. 1996 0.019 39
RG network 1996 0.012 23
Random net.2041 0.003 5
957 0.001 1
• The network similarity measure between secondary and primary networks are significantly higher than other network.
( MST : 90.8 % , percolation : 76.6 % )
The secondary networks well represent the primary network.
) otherwise ( 0
linked) are ji, (if 1 ijw
Conclusions & Future Works• Conclusions
– Two secondary networks, MST & percolation network, reproduce the scale-free behavior and its degree of each node is in proportion to degree of primary network. Its degree contains the global information of primary network.
– Similarity measurement shows that the secondary networks reproduce original network quite well. ( MST: 91% , percolation: 77% )
– BC change Δbi(j) values represent the interaction between i-node and j-node. And It is related to diameter change directly.
– Δbi(j) and b(i) relations might help to explain network classification with BC distribution exponents.
• Future Works
– BC change calculation for other network models and real networks.– Precise relation between Δbi(j) and b(i) , analytic calculation.– Finding primary network from secondary network information.
Distribution of BC Changes .)(kbi
)()()( jbjbjb oii
ij ij
oioiij
ii ibbbjbjbjbb )()()()(
lk
lko
o dNN
Dd ,)1(
1
o
i
o
oi
o
oiN
o
oi
o
oi
o
oii
b
ibb
b
bb
D
DD
D
DDN
N
NN
DNN
D
NN
D
d
dd
)( 2
)1(
)1()2)(1(
1
bi : summation of BC after i-th node removedbo : summation of BC over whole network.
)1()(
)1( ,
NNib
NNbbdD
i
lklk
bκ : summation of BC from κ-th node to all.
AB
CF
E
D
G
start2
1
2
1
4
1
4
1
2
1
Distribution of BC Changes .)(kbi
ib
Δbi : ( i-th node removed ) summation of BC changes.
Network deformation =
Lost a source of BC =
)1(22 Nb
select alternative shortest path + detour
( Contribution to Δbi < 0 )
( Contribution to Δbi > 0 )
select alternative shortest path
+
detour
bdd
ll ,,
lk lk lk
lklk g
igibib
,
,,
)()()(
)(,, igg lklk
)(,, igg lklk
Nonlinear!
Contribution of Δbi = portion of b(i)
77.4%
22.6%
1
2
Distribution of BC Changes .)(kbi
A
B
A Small closeness centrality of A Large sum of distance from A Large ② contribution and small network deformation
B
:
: Large closeness centrality of B small sum of distance from B small ② contribution and large network deformation
bdd
ll ,,
Network
0)1(22)1(22 , NdNbb AA
A
0)1(22)( , NdBbb BBoB ( λ : detour length )
1 10 100 10000.1
1
10
100
1000
10000
100000
1000000
Nd(k) ~ k -2.9
Num
ber
of N
odes
( N
d )
Degree (k)
Introduction : Scale-free network• What is the Scale Free Network?
– SF network is the network with the power-law degree distribution.
Ex) BA model growth and preferential attachment
A.-L.Barabási and R. Albert, Emergence of scaling in random networks, Science, 286, 509 (1999)
Ex) Empirical Results of Real Networks World-Wide Web, Internet, Movie actor collaboration network, Science collaboration graph, Cellular network, etc.
R. Albert, H. Jeong, and A.-L.Barabási, Nature(London), 406, 378 (2000)
kkD ~)(
0
21
34
5
6
Next One?
– SF network shows error and attack tolerance.
1000 10000 100000 1000000 1E7 1E81E-7
1E-6
1E-5
1E-4
1E-3
0.01
0.1
1
10
D (
)
( load by SHT )
Introduction : Load & Classification of networks
• What is the “Load” ?– When every pair of nodes in a network exchanges data packets
along the shortest path, load, or “betweenness centrality(BC),” of a node is the total number of data packets passing through that node.
Ex) Internet traffic jam, influential people in social network, etc.
A
B
C
F
E
D
G
start2
1
2
1
4
1
4
1
2
1
– “It is found that the load distribution follows a power-law with the exponent δ~2.2(1)”
K.-I. Goh, B. Kahng, and D. Kim, Universal Behavior of Load Distribution in Scale-Free Networks, PRL, 87, 27 (2001)
2.2~)( D
- The exponent of load is robust without network model dependency. It can be used to classify the networks.
Kwang-Il Goh, et al., Classification of scale-free networks, PNAS, 99, 20 (2002)
δ is universal value !