scalable methods for the analysis of network‐based data · p. smyth: networks muri meeng, june 3,...

P.Smyth:NetworksMURIMee5ng,June3,2011

1

Scalable Methods for the Analysis of Network‐Based Data

MURI Mee=ng, June 3rd 2011, UC Irvine

Principal Inves=gator: Professor Padhraic Smyth Department of Computer Science University of California, Irvine

Addi5onalprojectinforma5ononlineatwww.datalab.uci.edu/muri


2

Today’sMee5ng

• Goals– Reviewourresearchprogress– Discussion,ques5ons,interac5on– Feedbackfromvisitors

• Format– Introduc5on– Researchtalks

• Regular:20minutes+5minsatendforques5ons/discussion• Short:10minutes(sessionaVerlunch)

– Twoopendiscussionsessions,ledbyfaculty– Ques5on/discussionencouragedduringtalks– Severalbreaksfordiscussion

Butts


3

MURIProjectTimeline

• Ini5al3‐yearperiod– May12008toApril30th2011

– Fundingactuallyarrivedtouniversi5esinOct2008

• 2‐yearextension:– May12011toApril30th2013

• Mee5ngs(allatUCIrvine)– KickoffMee5ng,November2008– WorkingMee5ngs,April2009,August2009– AnnualReview,December2009– WorkingMee5ng,May2010

– AnnualReview,November2010– Today,June2011


4

Mo5va5on

2007:interdisciplinaryinterestinanalysisoflargenetworkdatasets

Manyoftheavailabletechniquesweredescrip5ve,couldnothandle

‐ Predic5on

‐ Missingdata

‐ Covariates,etc


5

Mo5va5on



‐ Predic5on

‐ Missingdata

‐ Covariates,etc

2007:significantsta5s5calbodyoftheoryavailableonnetworkmodeling

Manyoftheavailabletechniquesdidnotscaleuptolargedatasets,notwidelyknown/understood/used,etc


6

Mo5va5on



‐ Predic5on

‐ Missingdata

‐ Covariates,etc

2007:significantsta5s5calbodyoftheoryavailableonnetworkmodeling

Manyoftheavailabletechniquesdidnotscaleuptolargedatasets,notwidelyknown/understood/used,etc

Goal of this MURI project

Develop new sta=s=cal network models and algorithms to broaden their scope of applica=on to large, complex, dynamic

real‐world network data sets


7

MURITeam

Inves=gator University Department Exper=se Number Of PhD Students

Number of Postdocs

PadhraicSmyth(PI) UCIrvine ComputerScience Machinelearning 4

CarterBugs UCIrvine Sociology Sta5s5calsocialnetworkanalysis

6

MarkHandcock UCLA Sta5s5cs Sta5s5calsocialnetworkanalysis

1 1

DaveHunter PennState Sta5s5cs Computa5onalsta5s5cs

2 1

DavidEppstein UCIrvine ComputerScience Graphalgorithms 2

MichaelGoodrich UCIrvine ComputerScience Algorithmsanddatastructures

1 1

DaveMount UMaryland ComputerScience Algorithmsanddatastructures

2

TOTALS 18 3


8Collabora5onNetwork

Padhraic Smyth Dave

Hunter

Mark Handcock

Dave Mount

Mike Goodrich

David Eppstein Carter

Butts


9

Emma Spiro

Lorien Jasny

Zack Almquist

Chris Marcum

Sean Fitzhugh

Nicole Pierski

Collabora5onNetwork

Padhraic Smyth Dave

Hunter

Mark Handcock

Dave Mount

Mike Goodrich


Butts

Chris DuBois

Minkyoung Cho

Eunhui Park

Miruna Petrescu-Prahova

Arthur Asuncion Jimmy

Foulds

Duy Vu Ruth Hummel

Michael Schweinberger

Ranran Wang

Nick Navaroli

Darren Strash

Lowell Trott

Maarten Loffler

Joe Simon


10

Emma Spiro

Lorien Jasny

Zack Almquist

Chris Marcum

Sean Fitzhugh

Nicole Pierski

Collabora5onNetwork

Padhraic Smyth Dave

Hunter

Mark Handcock

Dave Mount

Mike Goodrich


Butts

Chris DuBois

Minkyoung Cho

Eunhui Park



Foulds

Duy Vu Ruth Hummel


Ranran Wang

Nick Navaroli

Darren Strash

Lowell Trott

Maarten Loffler

Joe Simon


11

Emma Spiro

Lorien Jasny

Zack Almquist

Chris Marcum

Sean Fitzhugh

Nicole Pierski

Collabora5onNetwork

Padhraic Smyth Dave

Hunter

Mark Handcock

Dave Mount

Mike Goodrich


Butts

Chris DuBois

FACEBOOK

Romain Thibaux

Minkyoung Cho

Eunhui Park



Foulds

Duy Vu Ruth Hummel


Ranran Wang

Nick Navaroli

Ryan Acton

Krista Gile

Computational Social Science Initiative U Mass Amherst

GOOGLE

INTEL

Darren Strash

Lowell Trott

Maarten Loffler

Joe Simon

RAND


12

Example:NetworkDynamicsinClassroomsNicole Pierski, Chris DuBois, Carter Butts


13

Data: Count matrix of 200,000 email messages among 3000 individuals over 3 months

Problem: Understand communication patterns and predict future communication activity

Challenges: sparse data, missing data, non-stationarity, unseen covariates


14

KeyScien5fic/TechnicalChallenges

• Parametrizemodelsinasensibleandcomputableway– Respecttheoriesofsocialbehavioraswellasexplainobserveddata

• Accountforrealdata– E.g.,understandsamplingmethods:accountformissing,error‐pronedata

• Makeinferencebothprincipledandprac5cal– computa5onally‐scalability:wantaccurateconclusions,butcan’twaitforeverfor

results

• Dealwithrichanddynamicdata– Real‐worldproblemsinvolvesystemswithcomplexcovariates(text,geography,

etc)thatchangeover5me

Insum:sta5s5callyprincipledmethodsthatrespectthereali5esofdataandcomputa5onalconstraints


15

DomainTheory DataCollec5on

Sta5s5calModels

MappingtheProjectTerrain


16


Sta5s5calModels Sta5s5calTheory



17

DataStructuresandAlgorithms



Es5ma5onAlgorithms



18

DataStructuresandAlgorithms



Es5ma5onAlgorithms

Inference

HypothesisTes5ng

Predic5on/Forecas5ng

DecisionSupport

Simula5on



19

Topic State of the Art Prior to Project

State of the Art Now

Poten=al Applica=ons And Impact

Generaltheoryforhandling

missingdatainsocialnetworks

Problemonlypar5allyunderstood.

NosoVwareavailableforsta5s5calmodeling

Generalsta5s5caltheoryfortrea5ngmissingdatainasocialnetworkcontext.

Publicly‐availablecodeinR.(GileandHandcock,2010)

Allowsapplica5onofsocialnetworkmodelingtodatasetswithsignificantmissing

data

Rela5onaleventmodels

Basicdyadiceventmodels.

Noexogenousevents.NopublicsoVware.

Muchrichermodelwithexogenousevents,egocentricsupport,mul5pleobserver

accounts,hierarchiesSoVwarepubliclyavailable

(Bugsetal,2010)

Providesageneralframeworkfordynamic

networkmodelingtolargerealis5capplica5ons

Cliquefindingalgorithms

Tooslowforuseinsta5s5calnetwork

modeling

Newlinear‐5mealgorithmforlis5ngallmaximalcliquesin

sparsegraphs(Eppstein,Loffler,Strash,

2010)

Extendsapplicabilityofsta5s5calnetworkmodelingtolargernetworksandmore

complexmodels

Accomplishments


20











data






(Bugsetal,2010)





modeling



2010)


complexmodels

Accomplishments


21











data






(Bugsetal,2010)





modeling



2010)


complexmodels

Accomplishments


22











data






(Bugsetal,2010)





modeling



2010)


complexmodels

Accomplishments


23Application to Email Data: 200,000 email messages among 3000 individuals over 3 months

(DuBois and Smyth, ACM SIGKDD 2010)


24

Impact:SoVware

• RLanguageandEnvironment– Open‐source,high‐levelenvironmentforsta5s5calcompu5ng

– Defaultstandardamongresearchsta5s5cians‐increasinglybeingadoptedbyothers

– Es5mated250kto1millionusers

• Statnet– Rlibrariesforanalysisofnetworkdata– Newcontribu5onsfromthisMURIproject:

• Missingdata(GileandHandcock,2010)

• Rela5onaleventmodels(Bugs,2010)

• Latent‐classmodels(DuBois,2010)

• Fastclique‐finding(Strash,2010)• +more……


25

Impact:Publica5ons

• Over40peer‐reviewedpublica5ons– acrosscomputerscience,sta5s5cs,andsocialscience

– Highvisibility• Science,Bugs,2009• Journal of the American Sta3s3cal Associa3on,Schweinberger,inpress• Annals of Applied Sta3s3cs,GileandHandcock,2010• Journal of the ACM,daFonsecaandMount,2010• Journal of Machine Learning Research,Asuncion,Smyth,etc,2009

– Highlyselec5veconferences• ACMSIGKDD2010(16%acceptrate)• NeuralInforma5onProcessing(NIPS)Conference2009(25%accepts)• IEEEInfocom2010(17.5%accepts)

• Cross‐pollina5on– Exposingcomputerscien5ststosta5s5calandsocialnetworkingideas

– Exposingsocialscien5stsandsta5s5cianstocomputa5onalmodelingideas


26

Impact:WorkshopsandInvitedTalks

• 2010Poli5calNetworksConference– WorkshoponNetworkAnalysis

– PresentedandrunbyBugsandstudentsSpiro,Fitzhugh,Almquist

• InvitedTalks:ConferencesandWorkshops– R!2010ConferenceatNIST(Handcock,2010)– 2010SummerSchoolonSocialNetworks(Bugs)– MiningandLearningwithGraphsWorkshop(Smyth,2010)– NSF/SFIWorkshoponSta5s5calMethodsfortheAnalysisofNetworkData(Handcock,2009)– Interna5onalWorkshoponGraph‐Theore5cMethodsinComputerScience(Eppstein,2009)

– Quan5ta5veMethodsinSocialScience(QMSS)Seminar,Dublin(Almquist.2010)– +manymore…..

• InvitedTalks:Universi5es– Stanford,UCLA,GeorgiaTech,UMass,Brown,etc


27

Impact:theNextGenera5on

• Facultyposi5onsatUMass– RyanActon,KristaGile‐>AsstProfs,partofnewini5a5veinComputa5onalSocial

Science

• Studentsspeakingatmajorsummerconferences– SunbeltInterna5onalSocialNetworks(Jasny,Spiro,Fitzhugh,Almquist,DuBois

– ACMSIGKDDConference(DuBois)

– Interna5onalConferenceonMachineLearning(Vu)

– AmericanSociologicalAssocia5onMee5ng(Marcum,Jasny,Spiro,Fitzhugh,Almquist)

• Bestpaperawardsornomina5ons(Spiro,Hummel)

• Na5onalfellowships:DuBois(NDSEG),Asuncion(NSF),Navaroli(NDSEG)


28

…..andtheOldGenera5on

• CarterBugs– AmericanSociologicalAssocia5on,LeoA.Goodmanaward,2010

– highestawardtoyoungmethodologicalresearchersinsocialscience

• MichaelGoodrich– ACMFellow,IEEEFellow,2009

• PadhraicSmyth– ACMSIGKDDInnova5onAward2009

– AAAIFellow2010

• MarkHandcock– FellowoftheAmericanSta5s5calAssocia5on,2009


29

WhatNext?

• “Push”algorithmicadvancesintosta5s5calmodeling– Willallowustoscaleexis5ngalgorithmstomuchlargerdatasets

• Developnetworkmodelswithricherrepresenta5onalpower– Geographicdata,temporalevents,textdata,actorcovariates,heterogeneity,etc

• Systema5callyevaluateandtestdifferentapproaches– evaluateabilityofmodelstopredictover5me,toimputemissingvalues,etc

• Applytheseapproachestohighvisibilityproblemsanddatasets– E.g.,onlinesocialinterac5onsuchasemail,Facebook,Twiger,blogs

• MakesoVwarepubliclyavailable


30

Organiza5onalCollabora5onduringtheKatrinaDisaster

Combined Network plotted by HQ Location

Almquist and Butts


31SESSION 1:

9:20 Dynamic Egocentric Models for Citation Networks Dave Hunter, Professor, Statistics, Penn State University 9:45 Membership Dimension Maarten Loffler, Postdoctoral Fellow, Computer Science, UC Irvine 10:05 Multilevel Network Models for Classroom Dynamics Chris DuBois, PhD student, Statistics, UC Irvine 10:30 COFFEE BREAK

SESSION 2: 10:45 DISCUSSION: FAST CHANGE-SCORE COMPUTATION IN DYNAMIC GRAPHS Led by David Eppstein and Michael Goodrich, Computer Science, UC Irvine 11:35 Bayesian Meta-Analysis of Network Data via Reference Quantiles Carter Butts, Professor, Social Sciences, UC Irvine

12:00 Break for lunch (lunch for PIs + visitors at the University Club)


32

1:30 to 2:40 Short Highlight Talks

Computational Issues with Exponential Random Graph Models Mark Handcock, Professor, Statistics, UCLA

Experimental Results on Fast Clique Finding David Eppstein, Professor, Computer Science, UC Irvine

Modeling Rates between Affiliates on Facebook from Sampled Data Emma Spiro, PhD student, Social Sciences, UC Irvine

Modeling Degree Sequences of Undirected Networks with Application to 9/11 Disaster Networks Miruna Petrescu-Prahova, Postdoctoral Fellow, Statistics, University of Washington

Statistical Models for Text and Networks Jimmy Foulds, PhD student, Computer Science, UC Irvine

Analysis of Life History Data Sean Fitzhugh, PhD student, Social Sciences, UC Irvine

Approximate Sampling for Binary Discrete Exponential Families with Fixed Execution Time and Quality Guarantees Carter Butts, Professor, Social Sciences, UC Irvine


33

2:40 – 3:30: SESSION 3

2:40 Instability, Sensitivity, and Degeneracy of Discrete Exponential Families Michael Schweinberger, Postdoctoral Fellow, Penn State University

3:05 Empirical Analysis of Latent Space Embedding David Mount, Professor, University of Maryland

3:30 COFFEE BREAK

4:00 DISCUSSION: LATENT VARIABLE MODELING OF NETWORK DATA Led by Carter Butts and Padhraic Smyth

4:45 WRAP-UP, CLOSING COMMENTS

5:00 ADJOURN 5:00 ADJOURN


34

Addi5onalResourcesProjectWebsite:hgp://www.datalab.uci.edu/muri/

SlidesandPostersfromAHM:hgp://www.datalab.uci.edu/muri/june2011/

Publica5ons:hgp://www.datalab.uci.edu/muri/publica5ons.php

SoVware:hgp://csde.washington.edu/statnet/

DataSets:hgp://networkdata.ics.uci.edu/resources.php


35

QUESTIONS?

scalable methods for the analysis of network‐based data · p. smyth: networks muri meeng, june 3,...

Documents