Missing Link Predition and Fuzzy
Communities
Ph.D. thesis booklet
Tamás Nepusz
Advisors:
Dr. FülöpBazsó(MTAKFKI-RMKI)
Dr.GyörgyStrausz(BME MIT)
BudapestUniversityofTehnologyandEonomis,
DepartmentofMeasurementandInformationSystems
2008
This dissertation is about networks:omplex systems onsist-
ing of uniqueelements onnetedby binaryrelations arranged
inseeminglyrandombutintrinsiallystruturedpatterns. Net-
worktheorywassuessfullyappliedtomodelvariousreal-world
phenomena ranging from the interations of proteinsin living
organisms to the large-saleorganization of human soiety or
to the struture of man-made tehnologial networks like the
Internet. Sinenetworks arebuiltfrom binaryrelationsamong
entities,theyannaturallybetransformedintographs,allowing
onetostudynetworkpropertiesbythetoolsofawell-established
eldofmathematis, namelygraphtheory.
Mydissertationonsistsoftwomajorparts,andthisisalsore-
etedin theontentsofthisthesisbooklet.Intherstpart,I
desribe and examineastohasti graph model where verties
of the graph are assigned to vertex types, and the onnetion
probability of any two verties depends solely on the types of
thevertiesinvolved.Myprimaryaimwas toapply thismodel
totheproblemofpreditingunknownonnetionsinanetwork
whose onnetionalstruture isknownonly partially. These-
ond part of my dissertation investigates a method that nds
densesubgraphs(modules,ommunities,lusters)in anetwork
undertheassumptionthatthesenetworkmodulesarenotwell-
separated and vertiesof the network may belong to multiple
ommunitiesatthesametime.
Prediting the probability of unknown onne-
tions in omplex networks
Introdution
Mostofthestate-of-the-arttoolsinnetworksieneassumethat
the onnetions of the network being studied are either om-
pletelyknown,or evenifthey arenot,the unertainties in the
in many ases, espeially when our knowledge about the on-
netionsstemsfromexperiments.This isaommonsenarioin
biologyandsoiology.Apeuliarexampleisthegraphmodelof
neuralonnetionsin the ortex,sinethe existeneofagiven
onnetion between two ortial areas an only be proved or
disprovedbyexpensiveandompliatedexperimentswithsome-
times ambiguous results.It is thereforeof primary importane
to estimate the probability of the existene of yet unharted
neural onnetions based on the known ones in order to help
experimenterstoonentrateonthosethat arelikelytoexist.
There aremultiple waysto taklethe link preditionproblem,
whihanroughlybelassiedasfollows:
Methodsbased on loal similarityindies. Thesemethods
alulateasimilaritymeasurebetweenallpairsofverties
inthegraphbasedonsimpleloalpropertiesoftheverties
(and possiblysomeadditionalapriori information). The
ommon underlying assumption ofthese methods is that
unharted onnetions are likely to exist between vertex
pairswithhighsimilaritysores.These soresareusually
derivedfromthesetofneighboursofthevertiesand the
amountofoverlapbetweenthesesets. Someexemplarsof
these methods are theJaard similaritysoreorthe in-
verse log-weighted similarity index of Adami and Adar
[1℄.
Methodsbased on paths and randomwalks. Thesemeth-
odsassessthesimilarityoftwogivenvertiesfromtheset
of paths or random walksbetweenthem. Most probably
oneoftheoldestmethodfromthisfamilyistheKatzsim-
ilarity sore [10℄, whih onsiders all the possible paths
betweenvertexpairs.Sine in generalthereare innitely
manypathsbetweenanytwooftheverties(allowingver-
tex repetitions in paths), the weight of longer paths is
dampedexponentiallyto keepthesumnite. When
P x,y hki
denotes the set of paths of length
k
between vertiesx
and
y
, and0 < β < 1
is the damping fator, the Katzsimilarityof
x
andy
isgivenby:∞
X
k=1
β k |P x,y hki |
(1)Anotherfrequentlyusedpath-basedsimilaritymeasureis
SimRank [8℄. Similarly to the well-known Google Page-
Rankmeasure,SimRankisdenedbyaself-onsistentre-
ursive equation. The basi ideais that two vertiesare
similar iftheyhaveinomingedges fromsimilar verties.
Let
Γ + (x)
bethesetofpredeessorvertiesofx
(i.e.,foreveryvertex in
Γ + (x)
, there exists at least a singleedgethatoriginatesfromthisvertexandterminatesin
x
),andlet
0 < γ < 1
anappropriatedampingfator (a ommon hoie is 0.8 [8℄). The denition of SimRank is then asfollows:
SimRank
(x, x) = 1
SimRank
(x, y) = γ P
a∈Γ + (x)
P
b∈Γ + (y)
SimRank(a, b)
|Γ + (x)| |Γ + (y)|
(2)
Itanbeshownthatthesolutionoftheequationsaboveis
uniqueandanbedeterminediteratively[8℄.Atthesame
time, SimRankis theexpetedvalueof
γ L,where L
is a
random variable desribing thenumberof steps required
fortworandomwalksstartedfrom
x
andy
tomeetforthersttime.
Aommondisadvantageofthemethodsdesribedaboveisthat
unknown or unertain onnetions are treatedas nonexistent,
thus they aremore appropriate forpredition problems where
onehas to extrapolateto the future behaviour of thenetwork
pletely[11℄.
Thereisalsoathird,relativelynewapproahofthelinkpredi-
tionproblem:letusrstonstrutanappropriaterandomgraph
modelthat issophistiatedenoughtodesribethenetworkbe-
ing studied, then nd the parameterisation of the model that
reproduesthe givennetwork withthehighestprobability,and
usetheprobabilityoftheexisteneofourunknownonnetions
inthemodeltoestimatetheirprobabilityintherealnetwork.If
theoriginalrandomgraphmodelisabletohandleknownnonex-
istingandunknownonnetionsdierently,oneanexpetmore
auratepreditionsfromthesemethodsthanfromtheonesthat
donotdistinguishbetweenthem. Themethoddesribedinmy
dissertationutilisestheseideas.
Researhgoals
Mygoalwasto oneiveamethod thatis abletoestimate the
probability of unknown onnetions in a given stati network
underthefollowingassumptions:
1. All thevertiesin thenetwork areknown,thepossibility
ofaddingnewvertiesormergingexistingonesshouldnot
beonsidered.
2. Vertexpairsareeitheronneted(andonrmed),dison-
neted(andonrmed),orunertain(theironnetedness
isunknown,butwemayhaveanassoiatedaprioridegree
ofbeliefregardingitsexistene).
3. Vertexpairs areordered; inother words, theonnetions
are direted, and the existene of a onnetion from
A
to
B
doesnotimply theexistene ofaonnetion in theoppositediretion.
Idenedastohastigraphmodel(alledthepreferenemodel),
where theonnetion probabilitiesof vertexpairsaregoverned
by vertex types assigned to the verties involved. The model
uses
k
types,andeveryvertexhastwotypesat thesametime.Letusallthese typesin-types andout-types.Thein- andout-
types of thevertiesare enoded by integersbetween 1and
k
in vetors
~u = [u i ]
and~v = [v i ]
. The model also ontains apreferenematrix
P = [p ij ]
ofsizek ×k
,wherep ij desribesthe
probabilitythat avertex with out-type
i
onnets to a vertexwithin-type
j
. Therefore,theregularityinthestrutureof thenetworkisdesribedbythetypeassignmentsofthevertiesand
thepreferenematrix.Fittingthisstohastigraphmodeltothe
networkbeingstudiedisthebasisoftheprobabilityestimation
outlinedin myresearhgoals.
Sine the network an ontain unertain onnetions, I ould
nothaveusedtraditional graphdesriptionssuh asadjaeny
matriesoradjaenylists.I hadto ndadesriptionthat ex-
tendsoneofthesedesriptionsin awaythatenablesustotake
intoonsiderationtheadditionalinformationregardingtheun-
ertainty and the degree of belief assoiated to eah possible
onnetion.Thenetworkbeingstudiedisdesribedbya
b ij be-
liefvalueforevery
(i, j)
orderedvertexpair:b ij = 1
intheaseofaonrmedexistingonnetion,
b ij = 0
foraonrmedmiss-ingonnetionand
0 < b ij < 1
forunertainonnetions,withhighervaluesorrespondingtohigherdegreesofbeliefinother
words, to higherapriori probabilitiesbasedonour additional
domain-speiknowledge.Thematrix
B = [b ij ]
equalsthead-jaeny matrixof theoriginal graphif itis ompletely known,
thuseetivelydemonstratingthattheadjaenymatrixrepre-
sentationissimplyaspeialaseofthebeliefmatrix.Thebelief
matrixanbestoredeientlyasasparsematrixifthenetwork
issparseandmostmissingonnetionsareonrmed.
Thepreferene modelan bettedto anarbitrarynetworkby
likelihood, given thebelief matrix
B
of thenetwork.The esti- matedaposteriori probabilitiesanthenbedeterminedbythettedvertextypesandtheirorrespondingentriesin thealu-
latedoptimalpreferenematrix.
Let
G 0denotethegraphbeinganalysedandletn
bethenumber
of verties in
G 0. The likelihood of a given parameterisation
θ = (k, ~u, ~v, P )
isthenasfollows:L(θ|G 0 ) =
n
Y
i=1 n
Y
j=1 j6=i
b ij p u i ,v j + (1 − b ij )(1 − p u i ,v j )
(3)
Pratialappliations usethelogarithmof thelikelihoodin or-
derto avoidroundingerrorsand numerialinstabilitiesaused
byoating pointalulatinginvolvingverysmall probabilities.
The log-likelihood an be maximised by one of the following
methodsortheirombination:
Expetation-maximization(EM) method. Startingfroma
randomtypeassignment,oneanndaloaloptimumby
repeatedlyapplyingtwosteps.Oneof themisthe E-step
(based onthe rstletter of expetation), while theother
oneisalledtheM-step(denotingmaximisation).Bytem-
porarily assuming
~u
and~v
(the type assignments) to be onstant,the E-stepdeterminesE (log L(θ|G 0 ))
andthenestimates
E p ij for every possible type pair. The M-step
usestheestimated
P
matrixresultingfrom theE-stepto modifythevertextypesinawaythatmaximisestheloalontributionofeveryvertextothelog-likelihoodunderthe
assumptionthat noothervertieswillhangetheirgroup
assignments. The algorithm stops when no modiation
wasperformedintheM-step(sinethisimpliesthatnoth-
ingwill hangeintheE-stepaswell).
Markov hain Monte Carlo (MCMC)method. Thisalgo-
rithmperformsarandomwalkin thespaeofallpossible
therandomwalkinvolveshangingthetypeassignmentof
asinglevertexhosenrandomly.Elementsofthe
P
prefer- enematrixarethenre-estimatedbasedonthenewong-urationsimilarlyto theE-stepin theEMalgorithm. The
newstateisaeptedunonditionallyasthenextstatein
therandomwalkifitslikelihoodishigherthanthelikeli-
hoodof theold state.If thenewlikelihood is lowerthan
the old one,the ratio of thenew and theold likelihoods
givestheprobabilityofaeptane.Thisshemeistheap-
pliationoftheMetropolisHastingsalgorithm[7℄forthis
spei problem, therefore the state probabilities in the
stationary distribution of the resultingMarkovhain are
proportionaltotheirlikelihoods.Bytakingalargenumber
ofsamplesfromthehainafter asuientlylongburn-in
period(whihletstheresidualeetsofthestartingstate
diminish), wean ndaparameterisation withhigh like-
lihood.
The shortomingof the EMalgorithm is that it anget stuk
in a loal maximum, but this is ounterbalaned by the fat
that it onvergesfast. The MCMC algorithm is free from this
shortoming, sine we oasionally allow steps towards worse
statesaswell.Inpratie,theadvantagesofthemethodsanbe
ombinedbyreplaingtheburn-instageoftheMCMCproess
by EM iterations. The Markov hain is then started from the
loalmaximumfoundbytheEMproess.
Results
T 1/1. I showed that vertex degreesin the networks gener-
atedbythepreferenemodelaredesribedbytheweightedsum
of Poisson-distributed random variables. I also proved a su-
ient onditionfortheexisteneof agiantomponentin these
networks.
rametersof the model to agiven network,taking into aount
thedegreesofbeliefassoiatedtothepossibleonnetionsinthe
network.I testedthevalidityofthese algorithmsonomputer-
generatedtestgraphs.
T 1/3. I showedthat theAkaike information riterion[2℄ is
abletohoosethemostappropriatenumberofvertexgroupsof
themodelinanunsupervisedmanner.
PubliationsrelatedtothesesT 1/1.,T 1/2.,T1/3.:
•
Nepusz T., BazsóF.: Likelihood-based lustering of di- reted graphs.In: IEEE Proeedings of the 3rd Interna-tional Symposium onComputational Intelligeneand In-
telligentInformatis,Agadir,Moroo,2830Marh2007,
pp.189194.
•
Nepusz T., Bazsó F.: Maximum-likelihood methods for data miningin datasetsrepresented by graphs.In:IEEEProeedingsofthe5thInternationalSymposiumonIntel-
ligent Systems and Informatis, Subotia, Serbia, 2425
August2007,pp.161-165.
•
Nepusz T.,Négyessy L.,Tusnády G., Bazsó F.:Reon-struting ortial networks: ase of direted graphswith
highlevelofreiproity.Toappearin:HandbookofLarge-
Sale Random Networks, editors: Béla Bollobás, Róbert
Kozma, Dezs® Miklós. Springer, 2008. ISBN 978-3-540-
69394-9.
Appliations
The appliability of the model and the algorithms is demon-
stratedintheeldofbiology,sinenetworkdatasetsinbiology
usually originate from experiments, therefore they frequently
ofthevisualandsensorimotorortiesofthemaaquemonkey
asdesribedin [12℄. This network inorporated45 brain areas
and 463 onrmed existing neural onnetions between them.
360pairsofareaswereknowntobedisonneted,andnoinfor-
mationwasavailableregardingtheremaining 1157pairs.Suh
unertainty poses a hallenge to even the state-of-the-art link
preditionapproahes.
The preferene model wasable to reonstrut the known part
of the ortial network with high ondene (92.7% of known
existentand83.1%ofknownnonexistentonnetionswerepre-
dited orretly). Resultspertaining to thevisual areas of the
networkdesribethemostexatreonstrutionpublishedinthe
literature so far, and the preditions regarding the unknown
onnetionsalso seemplausible in line withearlier reonstru-
tionattempts[5,9℄.ROCurveswerealsousedtoomparethe
methodto othergenerilink preditionmethods (seeFig.1).
Conlusion
I presented a method that is ableto estimate the probability
of unknownonnetionsin a stati,direted omplexnetwork,
taking domain-spei information into aount by the means
ofa priori onnetionprobabilities(belief values) andI shown
the appliability of the method on a real predition problem.
Themethodisrelevantnotonlyintheeldofbiology,butinall
problemswhereresearhersareonfrontedwithnetworkswhih
areknownto beinomplete.
Fuzzy ommunities in undireted networks
Introdution
Aommonfeature ofnetworksmodeling naturalphenomenais
sparseness: the vast majorityof possible onnetionsare miss-
link predition methods on theortial network dataset. AUC
= areaunder urve,attaining its maximum at 1when the re-
onstrutionisperfet.
ing,thusthenumberofatualedgesgrowslinearlyinthenum-
berofvertiesasthe network size tends to innity. Inspiteof
theirsparsity,thesenetworksfrequentlyontaindensesubmod-
ules, whih tend tooinide with largerfuntional units of the
network.Forinstane,densesubgraphsofasoialnetworkusu-
allyorrespondtoirlesoffriends,groupsofoworkersandso
on.Oneof themoststudied problemsof network theoryis the
eient identiation of suh dense subgraphs [4℄, also alled
modules, ommunities orlusters.It analsobedemonstrated
that these ommunities an overlap with eah other [14℄, but
mostommunitydetetionalgorithmsassumethateveryvertex
belongsto oneandonlyoneoftheommunities.Thedierene
betweentheoverlappingandthenonoverlappingapproahisil-
lustratedonFig.2.
examplegraph.Left:nonoverlappinglusteringwithtwolusters
aordingto thealgorithmof Clausetetal. [4℄.Right:overlap-
pinglusteringwithtwolusters.Thebridge-likepositionofthe
entralvertexisnotrevealedbythenonoverlappingapproah.
Researhgoals
Researhonalgorithmsthatareabletodetetoverlappingom-
munitiesisarelativelynewproblemin networksiene.Atthe
time whenI started myown investigations, there wasno algo-
rithmthat wasableto quantifyhowmuhdoesagivenvertex
belongtoagivenommunity;thealgorithmsavailablewereonly
abletodeidewhetheragivenvertexbelongstoagivenommu-
nityornot.Therefore,myaimwastodevelopanalgorithmthat
isabletoidentifyoverlappingommunitiesinomplexnetworks
and haraterise the membership degrees of the verties with
respet to the deteted ommunities. I assumed that edges in
thenetworkareundiretedandthenetworktopologyisknown
exatly.
Myresearhwasinspiredbythefuzzy
c
-meanslustering[3,6℄.Fuzzy
c
-means lustering wasprovento be usefuland eientin problemswhenthepointstobelusteredwereembedded in
an
n
-dimensional spae with anappropriate distane funtion.However,there isnosinglestraightforwardembeddinganddis-
tanefuntion forgraphs,soIhadtotakeadierentapproah.
Formally, the output of a fuzzy lustering is a fuzzy partition
matrix, denoted by
U = [u ki ]
from now on.u ki denotes the
membership degreeofvertex
i
in lusterk
. Thefollowingon-straintsareimposedontheelementsofthematrix:
1.
0 ≤ u ki ≤ 1
foralli
andk
.2.
0 < P n
i=1 u ki < n for all k
, where n
is the number of
vertiesinthegraph.Informally:lustersannotbeempty
and noluster anontainallthevertiesto thegreatest
possibleextent.
3.
P c
k=1 u ki = 1foralli
,where c
is thenumberoflusters.
Informally:thesumofallmembershipdegreespertaining
to agiven vertex is 1, thereforeweare not interested in
outlier vertiesthatdonotbelongtoanyofthelusters.
Myalgorithmisbasedonasimilarityfuntiondenedoverpairs
ofverties.Ifwethinkaboutthemembership degreesasprob-
abilities (
u ki is theprobability of the event that vertex i
is in
luster
k
), the probability of the event that vertiesi
andj
are in thesameluster equalsthe dotprodut of theirrespe-
tivemembershipvetors:
s ij = P c
k=1 u ki u kj
.Itfollowsthat thesimilaritymatrix
S = [s ij ]
basedontheseprobabilitiesissimplyS = U T U
.Thekeyassumptionofthealgorithmisthatthepres- eneof anedgebetweentwovertiesrelatesto theirsimilarity,while the absene of an edge implies dissimilarity. Therefore,
one should try to nd a matrix
U
that makesonneted ver- tiessimilaranddisonnetedvertiesdissimilar.Weandeneagoalfuntionthatquantiesthegoodnessoftforagiven
U
basedon the sumof squareddierenes betweenthe expeted
andtheatualsimilarity:
f ( U ) =
n
X
i=1 n
X
j=1
w ij (˜ s ij − s ij ) 2 ,
(4)where
w ij isanarbitraryweighingtermand˜ s ij = 1
ifandonlyif
i
arej
onnetedori = j
,zerootherwise.Thisgoalfuntionhas
tobeoptimisedwithrespettotheonstraintsdened above.
Theonstraintonthesumofmembershipdegreesofagivenver-
texanbeinorporatedintothegoalfuntionbyLagrangemul-
tipliers,leadingtoaonstrainednonlinearoptimisationproblem
where the individual variables (
u ki) an take values from the
range
[0; 1]
.Startingfrom arandomonguration,thevalueof thegoalfuntion an beoptimisedbystandardgradient-basedoptimisationmethods (e.g., steepest desent orthe method of
onjugategradients).
Oneoftheadvantagesoffuzzylusteringomparedtononover-
lapping lustering is that it is able to quantify the sharedness
of a vertex between groups. I introdued several measures to
ahievethatgoal:
Bridgeness. Intuitively,avertexis abridgebetweenommu-
nitiestothegreatestpossibleextentifitbelongstoallthe
lusters withthesamemembership degrees.Thisstateis
haraterised by amembership vetorwhose oordinates
are
1/c
.Theotherextreme iswhenthevertexbelongstoonly oneof the ommunities, resulting in a membership
vetorwith asingle element of 1 (all other elements are
zeros). Note that the varianeof the vetor omponents
is zero in the former ase and the maximal variane of
(c − 1)/c
is attained in the latter ase. The bridgenessmeasureanthereforebederivedfromthevarianeofthe
b i = 1 − v u u t
c c − 1
c
X
j=1
u ji − 1
c 2
(5)
Bridgeness an also be weighted by theentralityof the
vertex,allowingonetoltervertiesthatareoutliers(hav-
inglargebridgenesswithsmallentrality).
Exponentiatedentropy. Anotherpossibleapproahistoon-
siderthemembershipvetorofvertex
i
astheprobability massfuntionofadisreterandomvariableU iandalu-
latetheentropyofthevariable.Theentropyof
U i willbe
lowerifvertex
i
isanonoverlappingvertexandhigherifi
isasigniantoverlap.Byusingbase-2logarithmin
H(U i )
(theentropyof
U i),thenumberofsigniantommunities
anbeobtainedby
χ i = 2 H(U i ):
χ i = 2 − P c k=1 u ki log 2 u ki =
c
Y
k=1
u −u ki ki (6)
Results
T2/1. Idevisedandimplementedanalgorithmto ndfuzzy
ommunitiesinundiretednetworks.Thealgorithmisbasedon
themaximisationof aglobal goalfuntion derivedfrom vertex
similarities.I testedthevalidity ofthealgorithmonomputer-
generatedtestgraphs.
T2/2. IextendedthemodularitymeasureofNewman[13℄to
aount for the fuzziness of the obtained partitions. I showed
howanoneemploythe fuzziedmodularityto hoosethe op-
timalnumberof ommunities.
ommunitiesbyintroduingthebridgeness,theweightedbridge-
nessandtheexponentiatedentropymeasuresofthemembership
vetors.
PubliationsrelatedtothesesT 2/1.,T 2/2.,T2/3.:
•
Nepusz T., Petrózi A., Négyessy L., Bazsó F.: Fuzzyommunitiesandtheoneptofbridgenessinomplexnet-
works.PhysRevE,77(1):016107,2008.
•
Nepusz T.,Bazsó F., Strausz Gy.: Algorithmiidenti-ationofbridgevertiesinomplexnetworks.In:Proeed-
ings ofthe15
th
PhDMini-symposium, Budapest Univer-
sityofTehnologyandEonomis, pp.7881,2008.
Appliations
The appliabilityof the method is demonstrated again on the
ortialnetworkdatasetdesribedinthepreviouspart.Verties
ofthenetworkanbe lassiedasbrainareasrelatedto either
visualortatileinputproessing.Visualareasanalsobesub-
divided based on anatomial onsiderations. My expetations
were that fuzzy lusteringshould beableto ndthe bisetion
between visual and tatile input proessing areas and should
identify the areas related to the integration of visual and ta-
tile information as bridges(sine the integration task requires
strongonnetionsto bothlusters).
The most appropriate fuzzy lustering of the ortex was ob-
tained with four lusters. Two of these four lusters inluded
mostly visual areas, the remaining two ontained mostly ta-
tile input proessing areas. Only two areas were mislassied
andtheknownanatomialsubdivision ofthevisualortexwas
also reognisable.Therewere veareasthat were identied as
bridges based on the entrality-weighted bridgeness measure;
soresfortheseareasweresigniantlyhigherthantheaverage
ashigherlevelintegratoryareasaordingtoourpresentunder-
standingof thevisual andsensorimotororties. Thetwomis-
lassied areas were also among these ve bridges, suggesting
thatthelassiationerrorisausedbythebridge-likeposition
oftheseareas.
Conlusion
The methodology desribed in this part of the dissertation is
suitable for deteting ommunities in omplex networks even
whentheseommunitiesoverlaportheirboundariesarenotwell
dened.Somemoreillustrationsoftheresultsofthemethodon
real datasets are also provided. The deteted bridges deserve
furtherattention,sinethesevertiesmayplayaruialrolein
thesystemmodeledbythenetworkstruture.Furtherresearh
diretions inlude (butare notlimitedto)the extensionof the
method to direted and weighted graphs, outlier verties and
alternativesimilaritymeasures.
List of publiations
Publiationsrelated to the Ph.D. theses
1. Nepusz T.,Négyessy L.,Tusnády G., Bazsó F.:Reon-
struting ortial networks: ase of direted graphswith
highlevelofreiproity.Toappearin:HandbookofLarge-
Sale Random Networks, editors: Béla Bollobás, Róbert
Kozma, Dezs® Miklós. Springer, 2008. ISBN 978-3-540-
69394-9.
2. Nepusz T.,Bazsó F., Strausz Gy.: Algorithmiidenti-
ationofbridgevertiesinomplexnetworks.In:Proeed-
ings ofthe15
th
PhDMini-symposium, Budapest Univer-
sityofTehnologyandEonomis, pp.7881,2008.
ommunitiesandtheoneptofbridgenessinomplexnet-
works.PhysRevE,77(1):016107,2008.
4. Nepusz T., Bazsó F.: Maximum-likelihood methods for
data miningin datasetsrepresented by graphs.In:IEEE
Proeedingsofthe5thInternationalSymposiumonIntel-
ligent Systems and Informatis, Subotia, Serbia, 2425
August2007,pp.161-165.
5. Nepusz T., BazsóF.: Likelihood-based lustering of di-
reted graphs.In: IEEE Proeedings of the 3rd Interna-
tional Symposium onComputational Intelligeneand In-
telligentInformatis,Agadir,Marokkó,2830Marh2007,
pp.189-194.
Further related publiations
6. Négyessy L., Nepusz T., Zalányi L., Bazsó F.: Conver-
gene and divergene are mostly reiproated properties
of theonnetions in the network of ortial areas. Pro
RoySoLondonB,aeptedasis,2008.
7. NepuszT.,BazsóF.,StrauszGy.:Anewapproahforthe
lustering of direted graphs.In:Proeedings of the14
th
PhDMini-symposium,BudapestUniversityofTehnology
andEonomis, pp.3437,2007.
8. CsárdiG.,Nepusz T.:Theigraphsoftwarepakagefor
omplex network researh. InterJournal of Complex Sys-
tems 1695,2006.
9. PetróziA.,NepuszT.,BazsóF.:Measuringtie-strength
invirtualsoialnetworks.Connetions,27(2):4957,2006.
10. Négyessy L.,Nepusz T.,KosisL.,BazsóF.:Predition
ofthemainortialareasandonnetionsinvolvedin the
tatile funtion of the visual ortexby network analysis.
Eur JNeurosi,23(7):19191930,2006.
fortheanalysisoflargedatasetsrepresentedbygraphs.In:
Proeedings of the13
th
PhDMini-symposium, Budapest
UniversityofTehnologyandEonomis,pp.5253,2005.
Other publiations
12. PetróziA.,AidmanE.V.,NepuszT.:Capturingdoping
attitudes by self-report delarations and impliit assess-
ment: amethodology study. Substane Abuse Treatment,
PreventionandPoliy,3:9, 2008.
13. KovásZs.,NepuszT.,HamarG.:Egyszer¶mozgásanal-
izátorneurológiaihasználatra(inHungarian).Kórház-és
orvostehnika,43(1):3-8,2005.
14. Herzeg E.,BaumgartnerI.,FazekasG.,KovásZs.,Ne-
pusz T.: Mozgásanalizátor alkalmazása féloldali bénult
betegek kézfunkiójának felmérésére (in Hungarian). Re-
habilitáió14(4): 31-34,2004.
Bibliography
[1℄ L.A. Adamiand E.Adar. Friends andneighbors on the
Web. SoialNetworks,25(3):211230,2003.
[2℄ H. Akaike. A new look at the statistial model identi-
ation. IEEE Transations on Automati Control, 19(6):
716723,1974.
[3℄ J. C. Bezdek. Pattern Reognition with Fuzzy Objetive
FuntionAlgorithms.Plenum,NewYork,USA,1981.ISBN
978-0306406713.
[4℄ A.Clauset,M.E.J.Newman,andC.Moore.Findingom-
munity struture in verylargenetworks. Physial Review
E,70:066111,2004.
[5℄ L.daFontouraCosta,M.Kaiser,andC.C.Hilgetag. Pre-
ditingtheonnetivityof primateortial networks from
ology,1(16),2007.
[6℄ J.C.Dunn. A fuzzyrelativeoftheISODATAproessand
itsuseindetetingompatwell-separatedlusters. Jour-
nalofCybernetis, 3:3257,1973.
[7℄ W. K. Hastings. Monte Carlo sampling methods using
Markovhains and their appliations. Biometrika, 57(1):
97109,1970. doi:10.2307/2334940.
[8℄ G.JehandJ.Widom. SimRank:Ameasureofstrutural-
ontext similarity. In Proeedings of the Eighth ACM
SIGKDD International Conferene on Knowledge Disov-
eryandDataMining,pages538543,NewYork,2002.As-
soiationof ComputingMahinery.
[9℄ B.Jouve,P. Rosenstiehl, andM.Imbert. A mathematial
approah to the onnetivity between the ortial visual
areasofthemaaquemonkey.CerebralCortex,8(1):2839,
1998.
[10℄ L.Katz.Anewstatusindexderivedfromsoiometrianal-
ysis. Psyhometrika, 18(1):3943,1953.
[11℄ D.Liben-NowellandJ.Kleinberg.Thelinkpreditionprob-
lem forsoial networks. In Pro. 12th International Con-
fereneonInformationandKnowledgeManagement,2003.
[12℄ L.Négyessy, T.Nepusz, L.Kosis,and F. Bazsó. Predi-
tionofthemainortial areasandonnetionsinvolvedin
thetatilefutionofthevisualortexbynetworkanalysis.
EuropeanJournalof Neurosiene,23(2):19191930,2006.
[13℄ M.E.J.Newman. Fastalgorithmfordetetingommunity
strutureinnetworks.Physial ReviewE,69:066133,2004.
[14℄ G.Palla,I.Derényi,I.Farkas,and T.Visek. Unovering
theoverlappingommunitystrutureof omplexnetworks
innature andsoiety. Nature,435(7043):814818,2005.