Você está na página 1de 43

ower Laws and Scale lree neLs

(More + Lab)
LyLan Adar
Sl308/708 CMLx608
CcLober 3 2011
Some slldes from L Adamlc and C Mel
lf you haven'L yeL
W uownload Lhey pyLhon dlsLrlbuLlon from
hLLp//enLhoughLcom/
LasL Llme
W ower law dlsLrlbuLlons
W llLLlng
W whaL klnds of processes generaLe power laws?
W 8arabaslAlberL model for scalefree graphs
owerlaw ulsLrlbuLlon Lhe Long 1alls
W Many clLles wlLh a small
populaLlon
W A few clLles wlLh a very
large populaLlon
W hlgh skew (asymmeLry)
W Comparlng Lo a normal
dlsLrlbuLlon
Many people wlLh a medlaLe
helghL
AlmosL nobody wlLh a very
hlgh or very low helghL
symmeLry
opulaLlon of clLy
ercenLage of clLy
owerlaw dlsLrlbuLlon
llnear scale
loglog scale
hlgh skew (asymmeLry)
sLralghL llne on a loglog ploL
JhaL goes wrong wlLh sLralghLforward blnnlng
W nolse ln Lhe Lall skews Lhe regresslon resulL
10
0
10
1
10
2
10
3
10
1
10
0
10
1
10
2
10
3
10
1
10
5
10



dal a
= 1. l|l
have many more blns here
have few blns
here
llLLlng vla regresslon Lo Lhe cumulaLlve dlsLrlbuLlon
W flLLed exponenL (243) much closer Lo acLual (23)
10
0
10
1
10
2
10
3
10
1
10
0
10
1
10
2
10
3
10
1
10
5
10

x
l
r
e
q
u
e
r
c
y

s
a
r
p
|
e

>

x


dal a
-1 = 1. 13 l|l
Maxlmum llkellhood flLLlng besL
W ?ou have Lo be sure you have a powerlaw dlsLrlbuLlon (Lhls
wlll [usL glve you an exponenL buL noL a goodness of flL)
1
1
min
ln 1

+ =

3

3
W

are all your daLapolnLs and you have of Lhem


W for our daLa seL we geL 2303 preLLy close!
1oday
W Jrapup
8arabaslAlberL model for scalefree graphs
8ack Lo neLworks
growlng neLworks powerlaws
andom graph model
W 1he degree dlsLrlbuLlon ls glven by
colnfllps Lo see how many people you'll be connecLed Lo one coln fllp
per each of Lhe (n1) oLher nodes
probablllLy of connecLlng
3

5 5

3
5

'
+

'


=
1
) 1 (
1
8lnomlal
olsson
normal
llmlL small
llmlL large

0
5

=
SlmplesL random neLwork
W Lrdosenyl random graph each palr of nodes ls equally llkely
Lo be connecLed wlLh probablllLy p
W p 2*L/n/(n1)
W olsson degree dlsLrlbuLlon ls narrowly dlsLrlbuLed around k
p*(n1)

9
%

olsson degree dlsLrlbuLlon


ercolaLlon Lhreshold ln Lrdosenyl Craphs
average degree
s
l
z
e

o
f

g
l
a
n
L

c
o
m
p
o
n
e
n
L
av deg 099 av deg 118
av deg 396
ercolaLlon Lhreshold how many edges need
Lo be added before Lhe glanL componenL
appears?
As Lhe average degree lncreases Lo z 1 a
glanL componenL suddenly appears
eal Jorld neLworks are CfLen ower
Law 1hough
14
Cont|nenta| A|r||ne Source http]]wwwa|r||neroutemapscom]
13
SCENCE CTATON NDEX
( 3)
Nodes: papers Links:
clLaLlons
(S edner 1998)
(k) k

648
23
PL SLanley
1736 L papers (1988)
Cltatlon Networks
Puman lnLeracLlon neLwork
16
ual f naLure 2003
SLelze f Cell 2003
1he lnLerneL
17
Craph from AlberLLszl 8arabsl' s SlCl09 keynoLe
(lalouLsos lalouLsos and lalouLsos 1999)
Swedlsh sexweb
nodes people (lemales Males)
Llnks sexual relaLlonshlps
W lL's a small world
W MosL lndlvlduals reporL 12 parLners ln Lhe pasL 12 monLhs buL some
Source 1he web of human sexual conLacLs Lll[eroseL alnaLure 411 907908(21 !une 2001)
eal world neLworks are
ofLen power law Lhough
W Sexual neLworks
W MosL lndlvlduals reporL
12 parLners ln Lhe pasL
12 monLhs buL some
Source 1he web of human sexual conLacLs Lll[eroseL alnaLure 411 907908(21 !une 2001)
referenLlal ALLachmenL ln neLworks
W llrsL consldered by rlce 63 as a model for clLaLlon neLworks
each new paper ls generaLed wlLh m clLaLlons (mean)
new papers clLe prevlous papers wlLh probablllLy proporLlonal Lo Lhelr
lndegree (clLaLlons)
whaL abouL papers wlLhouL any clLaLlons?
W each paper ls consldered Lo have a defaulL clLaLlon
W probablllLy of clLlng a paper wlLh degree k proporLlonal Lo k+1
W ower law wlLh exponenL 2+1/m
8arabaslAlberL model
W undlrecLed model each node connecLs Lo
oLher nodes wlLh probablllLy proporLlonal Lo
Lhelr degree
Lhe process sLarLs wlLh some lnlLlal subgraph
each node comes wlLh m edges
W esulLs ln powerlaw wlLh exponenL 3
8aslc 8Amodel
W very slmple algorlLhm Lo lmplemenL
sLarL wlLh an lnlLlal seL of m
0
fully connecLed nodes
W eg m
0
3
now add new verLlces one by one each one wlLh exacLly m edges
each new edge connecLs Lo an exlsLlng verLex ln proporLlon Lo Lhe
number of edges LhaL verLex already has preferentio/ ottochment
easlesL lf you keep Lrack of edge endpolnLs ln one large array and
selecL an elemenL from Lhls array aL random
W Lhe probablllLy of selecLlng any one verLex wlll be proporLlonal Lo Lhe
number of Llmes lL appears ln Lhe array whlch corresponds Lo lLs degree
1 2
3
1 1 2 2 2 3 3 4 3 6 6 7 8
generaLlng 8A graphs conL'd
W 1o sLarL each verLex has an
equal number of edges (2)
Lhe probablllLy of chooslng any
verLex ls 1/3
W Je add a new verLex and lL wlll
have m edges here Lake m2
draw 2 random elemenLs from
Lhe array suppose Lhey are 2
and 3
W now Lhe probablllLles of
selecLlng 123or 4 are
1/3 3/10 3/10 1/3
W Add a new verLex draw a verLex
for lL Lo connecL from Lhe array
eLc
1 2
3
1 1 2 2 3 3
1 2
3
1 1 2 2 2 3 3 3 4 4
4
1 2
3
4
1 1 2 2 2 3 3 3 3 4 4 4 3 3
3
roperLles of Lhe 8A graph
W 1he dlsLrlbuLlon ls scale free wlLh exponenL 3
(k) 2 m
2
/k
3
W 1he graph ls connecLed
Lvery verLex ls born wlLh a llnk (m 1) or several llnks (m 1)
lL connecLs Lo older verLlces whlch are parL of Lhe glanL componenL
W 1he older are rlcher
nodes accumulaLe llnks as Llme goes on
preferenLlal aLLachmenL wlll prefer wealLhler nodes who Lend Lo be
older and had a head sLarL
verLex lnLroduced aL Llme L3
verLex lnLroduced aL Llme L93
1lme evoluLlon of Lhe connecLlvlLy of a verLex ln Lhe 8A model
W ?ounger verLex does noL sLand a chance
W aL L93 older verLex has 20 edges and younger verLex ls sLarLlng ouL wlLh 3
W aL L 10000 older verLex has 200 edges and younger verLex has 30
Source 8arabas| and A|bert Lmergence of sca||ng |n random networks Sc|ence 1999
LhoughLs
W 8A neLworks are noL clusLered
Can you Lhlnk of a growLh model of havlng
preferenLlal aLLachmenL and clusLerlng aL Lhe
same Llme?
W JhaL would Lhe neLwork look llke lf nodes are
added over Llme buL noL aLLached
preferenLlally?
W JhaL oLher processes mlghL glve rlse Lo power
law neLworks?
wrap up
W power law dlsLrlbuLlons are everywhere
W Lhere are good and bad ways of flLLlng Lhem
W some dlsLrlbuLlons are noL powerlaw
W preferenLlal aLLachmenL leads Lo power law
neLworks
W buL lL's noL Lhe whole sLory and noL Lhe
only way of generaLlng Lhem
Lab
generaLlng scale free neLwork wlLh a[ek
generaLlng scale free neLworks wlLh neLLogo
CeneraLlng scale free neLworks wlLh
a[ek
W 1wo general opLlons
Scale free
W uM ennock eL al (2002) Jlnners don'L Lake all nAS 99/8 3207
3211
W a[ek command -et kandom -etwork Sca|e Iree
W ulffers from Lhe 8A model prlmarlly ln LhaL
new verLlces are noL auLomaLlcally asslgned edges
probablllLy of aLLachlng ls parLlally lndependenL of degree
LxLended model
W AlberL 8arabasl AL 1opology of evolvlng neLworks local evenLs
and unlversallLy http]]xxx|an|gov]abs]condmat]000S08S
W a[ek command -et kandom -etwork Lxtended Mode|
W ulffers from Lhe slmple 8A model ln LhaL
edges are added beLween exlsLlng nodes noL only Lhe newcomer
edges are rewlred beLween exlsLlng nodes
flLs Lo varlous subseLs of web daLa
and web pages ln general
ennock model
W Lxample lL ls reasonable Lo assume LhaL some webpages wlll be llnked Lo
ln parL because of whaL Lhey are raLher Lhan Lhe number of llnks Lhey
already have
ennock model
W 1he probablllLy LhaL an endpolnL of a new edge
connecLs Lo verLex ls
9 2 29

+
+ = H

1
) 1 (
2
) (
#Scale free' ln a[ek
W lor Lhe neLwork you can speclfy
#undlrecLed' #dlrecLed' or #acycllc'
W an #addlng free' opLlon?
# of verLlces
# of llnes
average degree of verLlces
lnlLlal Lrdosenyl Craph (Lhese are Lhe flrsL few verLlces presenL)
W # of verLlces (use someLhlng small a couple of verLlces)
W probablllLy p of connecLlng Lype 09999 Lo have Lhem fully connecLed or anyLhlng
beLween 0 and 1 doesn'L maLLer much
Lhls ls beLween 0 and 03 for an undlrecLed graph
W Lhe hlgher Lhe more scalefree your dlsLrlbuLlon wlll be
W buL waLch ouL lf you seL 03 Lhen 03 and 0 and your new edgeless
verLlces wlll never geL new connecLlons you wlll only have Lhe orlglnal Lrdos
enyl componenL connecLed
ln Lheory you can leave elLher Lhe # of
verLlces or # of llnes unconsLralned buL
leavlng Lhe # of llnes unconsLralned
(enLer ln #0') works for me
#Scale free' neLwork opLlon ln a[ek
W neLwork sLarLs wlLh

verLlces whlch llnk Lo each oLher


wlLh probablllLy

(as ln an Lrdosenyl random graph)


W AL each Llme sLep one verLex and edges are added Lo
Lhe neLwork
W lnsLead of aLLachlng one end polnL of each edge Lo Lhe
newly lnLroduced verLex choose end polnLs accordlng
Lo Lhe probablllLy
'
;

;
;
4:9 3
1 ) ( deg ) ( deg
) Pr( + + =
fracLlon of edges ln Lhe
graph LhaL sLarL aL v
fracLlon of edges ln Lhe
graph LhaL end aL v
Lhe credlL v
geLs [usL for belng
one of Lhe verLlces
#Scale free' neLwork generaLlon ln
a[ekconL'd
W CbservaLlons
+ + 1 so can vary Lhe relaLlve lmporLance of lndegree ouLdegree
and lndependenL probablllLy
ln an undlrecLed neLwork = , slnce lndegree and ouLdegree are Lhe
same
noL all verLlces wlll be connecLed slnce Lhey are noL #born' wlLh an edge
1he larger ls Lhe less scalefree Lhe degree dlsLrlbuLlon
W edges are added wlLhouL regard Lo degree
W Crlglnal 8A paper showed LhaL ln LhaL case Lhe degree dlsLrlbuLlon
(k) exp(k) an exponenLlal dlsLrlbuLlon
'
;

;
;
4:9 3
1 ) ( deg ) ( deg
) Pr( + + =
LxLended 8A model (undlrecLed
neLwork)
W sLarL wlLh m
0
lsolaLed nodes
W aL each LlmesLep perform one of Lhe followlng operaLlons
w/ prob add m (m m
0
) new llnks
W for each llnk
selecL #from' verLex aL random
selecL #Lo' verLex ln proporLlon Lo lLs degree (+1 so LhaL lsolaLed verLlces have a chance of
geLLlng llnks)
w/ prob q where 0 q 1 p
W rewlre m llnks
selecL node l aL random and one of l's llnks
rewlre Lhe endpolnL of l's llnk Lo anoLher node [ randomly chosen wlLh probablllLy H(k
[
)

+
+
= H

1
1
) (
LxLended 8A model conL'd
w/ prob 1 p q
W add a new node wlLh m llnks
connecL endpolnLs of Lhe m llnks Lo verLlces ln proporLlon Lo Lhelr
degree (H(k
[
)
W ln Lhe pq0 llmlL reduces Lo Lhe slmple 8A
model
W ln Lhe hlgh q (q 1) llmlL exLended model
produces a neLwork wlLh an exponenLlal Lall
because growLh ls very slow (only rewlrlng ls
occurrlng)
parameLer space of Lhe exLended 8A
model
W ln Lhe hlgh p (p 03) llmlL have a scale free dlsLrlbuLlon because
addlng new edges preferenLlally
saLuraLlon effecL for small k (degree)
W because edges keep belng added buL verLlces are noL belng added LhaL
qulckly evenLually even Lhe low degree verLlces geL a few more edges
powerlaw exponenL varles beLween 2 and C dependlng on
parameLers
LxLended 8A model ln a[ek
W neL andom neLwork LxLended Model
W Speclfy
n # of verLlces
m
0
# of lnlLlal dlsconnecLed nodes
m m
0
number of edges Lo add/rewlre aL a Llme
p probablllLy Lo add new llnes
q probablllLy Lo rewlre edges 0 q 1p
can ask for neLwork wlLhouL mulLlple llnes
Craph randomlzaLlon
Pow can we randomlze a neLwork whlle
preservlng Lhe degree dlsLrlbuLlon?
W SLub reconnecLlon algorlLhm (M L newman eL al 2001 also known ln
maLhemaLlcal llLeraLure slnce 1960s)
W 8reak every edge ln Lwo edge sLubs"
APF8 Lo AF F8
W andomly reconnecL sLubs
W roblems
Leads Lo mulLlple edges
CannoL be modlfled Lo preserve addlLlonal Lopologlcal
properLles
Local rewlrlng algorlLhm
W andomly selecL and rewlre Lwo edges (Maslov Sneppen 2002 also known ln
maLhemaLlcal llLeraLure slnce 1960s)
W epeaL many Llmes
W reserves boLh Lhe number of upsLream and downsLream nelghbors of
each node (le lndegree and ouLdegree)
Conservlng addlLlonal lowlevel
Lopologlcal properLles
W ln addlLlon Lo k
l
one may also conserve
1he exacL numbers of loops or oLher moLlfs
1he slze and numbers of componenLs lnLerneL all
nodes have Lo be connecLed Lo each oLher
W MeLropolls algorlLhm Lwo edges are rewlred based on
L(n
acLual
n
deslred
)
2
/n
deslred
lf LA0 rewlrlng sLep ls always accepLed
lf L0 rewlrlng sLep ls accepLed wlLh pexp(L/1)
number of a cerLaln submoLlf you wanL Lo preserve

Você também pode gostar