0-云計(jì)算時(shí)代的社交網(wǎng)絡(luò)-平臺(tái)和技術(shù)-張智威

云

計(jì)

算

時(shí)

代

的

社

交

網(wǎng)

絡(luò)

平

臺(tái)

和

技

術(shù)張智威副院長(zhǎng)，研究院，谷歌中國(guó)教授，電機(jī)工程系，加州大學(xué)2/16/2011

Ed

Chang

1ChinaOpportunityChina

&

US

in

2006-07180

million

208

million60

million

60million

500

million

180million600

k72

kChinaU.S.2Mobile

PhonesEngineering

Graduates2/16/2011(125%)(13%)(190%)

(129%)InternetPopulationBroadband

UsersEd

Chang·Size(~700)-200

engineers-400other

employees—Almost

100internsLocations-Beijing

(2005)—Taipei(2006)—Shanghai(2007)Google

China2/16/2011●OrganizingtheWorld'sInformation,Socially·社區(qū)平

臺(tái)

(SocialPlatform)·云運(yùn)

算

(Cloud

Computing)·結(jié)論與前瞻(Concluding

Remarks)2/16/2011Ed

Chang40

□2/16/2011Ed

Chang

5-jpg.htmO.jpgLmsgWeb

1.0htmhtm.htm.htm.htm.docO

OWebwith

People(2.0)doc2/16/2011htmEd

Chang-jpgO

Omsg,.xls.htm.htm.htm-jpgmsgO

oO

Oa

e6.msg.xls.htm-jpgO-jpgOO.htmApp(Gadget)

.doc+Social

Platformspp(GaeEdChang2/16/2011.htmmsg.htmG

0ae7oDoneaddapplicationadd

applicationEd

Chang盤prafile

edit?

scrapboolphotosvideos○

teatimoniolsask

fnendsApps

editMusic

iLike面HcroscopesFunWall

by

Sl..listsmessagesupdatessettingsstartFlixster

Movies

食食☆☆☆addapolicatien

e

o

k

s

i

a

ppesli

ba

oik

e

riiRlilRc

d

o

us

nen

ye

aadvikl

ao

nooiounr★lv女esmicayrsh☆eadinigtorruwaeszaaccnmcanoveaucroisanmokgeotacewoacPeeapplicationdirectoryHome>EdChang>applicationdirectory<previous

I

next>

h

t

o

o

tA

t

h

e

Ir

i

o

ioo

sa

aapnt

mim

tililoy

hnadt

mh

l

i

a

o

eTw

pri

eetsn

fisk

sbt

i

ol

mpetitive

typing

game.Compete

against

yourfriends

and

thew合h由o女le音☆addapplicatiengcnroar-evefiort

oin

thew

sraoninmaclndgeRscool有南urthprassyothemwixhasolu,sosyspkoacanthen,sasneahve!5laswatarenimnol1ewaracwurnaos

ec

a

t

o

i

te

with

friends.Create

and

challenge

friends

to

movieesasttatervaoimilr

msuwirespanmfoidvnmos

ahertingotraMeetmovieeezrzauihqSDookmarks

Iools

LelpO

htp://AcoDrectory.asoxMG-In

…

EectricalaCom

… arkut

applicatian

diractory

Mozilla

Firafax

Horoscopes

山出擊合臺(tái)Login

G-C..echangP

1ogohGetyourhoroscopes

-Updated

every

other

day.MGooge,com-e

…QCdetdit

Yew

HttoryH

inenetfewsR…orkutDorkut-applic-|Flixster2ht

-scrkot.-?E"E*Pteet32eneL?*VTransferninodatafromlstbe.tanva.

cn

…star

whdbwnEtlere

PNeP@

oocd-c?d.

htee//ebedenyeE0m回

區(qū)資用友望入來(lái)吧郵件姓名發(fā)送道讀我要基發(fā)逐語(yǔ)》eee-eeeee-MozillaFiretexCle

Edit

Yew

Htgtory

Bookmarks

我的主頁(yè)

資料

圍友我的朋友圈|我的擁友wusmIoob

Eep來(lái)筆

站子PmnCagem

mas禮瘤

得價(jià)

蛋縣薄影集日記uanehaeg回到自己(0e●烏三巴托蒙古●呼和潔料●大西■濟(jì)南，鄭合肥哈爾濱長(zhǎng)春陽(yáng)朝鮮平壤首爾韓國(guó)日本0示京回

區(qū)MGoode

.com-Irbox(49)-edchang@g

…△成都●重莊費(fèi)陽(yáng)●星明武漢

南昌福州

約魚(yú)島●臺(tái)北拉薩不舟孟加拉國(guó)緬卸●柳光可富汗

●伊斯蘭堡巴基斯坦●新語(yǔ)里

尼泊名mee

MoyillaFiretoxCleEdit

Yew

Hgtory

Bookmsrks

Iook

8。天準(zhǔn)來(lái)吧·我的朋友型MoPeEPonoo00--t://see

srye.e.

-104●網(wǎng)斯培納

業(yè)

言爾吉斯斯組塔吉克斯坦、老過(guò)越南西沙群島G

S供羅斯聯(lián)邦■南速門NS●河內(nèi)

不砂桿應(yīng)Hep號(hào)。天涯來(lái)吧-蘋津完地圖數(shù)璃@2007

Mm

sta天涯來(lái)E-calo_qu的個(gè)人資料2WheeweEtlere●烏毒木齊克新●西安Done●●印度●●●i置化縣還花因區(qū)涿底縣—懷來(lái)縣aig

主于滿塊自治星共有20名用戶在比區(qū)域(第1頁(yè))共5頁(yè))eyuch

老湯翻火點(diǎn)擊查看細(xì)節(jié)信縣縣承

德

縣意昌

平區(qū)大興區(qū)順交區(qū)

平谷區(qū)

Q

遷西縣大廠回族自演縣寶坻是右照程尚

義

縣丹和縣天

鎮(zhèn)

縣陽(yáng)高

縣張北縣“懷安縣MGcodle.co

…8

天涯*-

8

天遵來(lái)吧…

8o

天蓮來(lái)吧

…Goole,co

…

8o

天通來(lái)吧…d

OpenSoa

…

C]Goode地…C]Develoer

…MGoode.co

…人ede

tdit

Yew

Htgtory

Bookmarke

Iook

Hepe

http:/Laba.tanya.onAalba/FiendMap?d=14914947603760770386其丘縣易縣武

清

縣地重稱柳意00g

spabccog-薛

縣廣民縣渾

源

縣天涯來(lái)吧MozillaFirefex深水縣高碎店市市太同

縣陽(yáng)原縣王田縣隆化縣mg承

潔

縣DoaCa下一頁(yè)5

翼

后Done○netbt區(qū)2/16/2011

Ed

Chang

12開(kāi)

放

社

區(qū)

平

臺(tái)Linked

inFriends

rciet

siesvilorkut

Linked

in

hi5

sale

force.comorkut

Linked

hi5

sale

force.comorkutLinked

inhi5面sale

force.comOpenSocial開(kāi)

放

社

區(qū)

平

臺(tái)我是誰(shuí)2/16/2011Ed

Chang

17社區(qū)平臺(tái)

我

的

朋

友他的活動(dòng)Fle

Edit

yewHytoryBookmarksIookLtep ho:Aabs

tnva

cnAsea

FustrtendsMGoode

.com-In

…

G]ooe

gadget

…

a

OperSoosl

Rat我的主而資料

朋友

采吧

帖子

彩集

日記我的朋友圈|我的朋友上一步1郵件發(fā)送造請(qǐng)我要群發(fā)激請(qǐng)。香看我的期2*e-CRP?E天涯來(lái)吧我的朋友圖-MorillaFiretexTransferrino

data

from

lstbe.tenva.cn

…天涯來(lái)吧-我_留意薄Feui

Dashbord禮物Pam..Darren

Hiang-

…Googie.com-C

…8評(píng)價(jià)年常帶用|xPaiment..園到自己star3理hdA天涯來(lái)吧我的朋友圖-MorillaFiretexFle

Edit

yewHytoryBookmarksIookLep·

ho:Aabs

tnva

cnAsea

FustrtendsMGoode

.com-In

…

G]ooe

gadget

…[C]我的主而資料

朋友

采吧

帖子

彩集日記我的朋友圈|我的朋友上一步1品天源*吧-載-評(píng)價(jià)

智音清Darren

Hiang-

…G-C

…A郵件發(fā)送造語(yǔ)我要群發(fā)激請(qǐng)。園到自己五最選的，選的然是105-1過(guò)直進(jìn)QQQ91

五公司的執(zhí)

扣2白級(jí)的建直求金101-12-30白級(jí)的實(shí)班家金101-1230白領(lǐng)的家庭重金101-1230移殊好友hel人·

永遠(yuǎn)來(lái)吧(離線)擔(dān)量上線想片吧1男

3

7

歲

北

京項(xiàng)口和?也

Pam..

Peaiment..

2*oERstarTransferrino

data

from

lstbe.tenva.cn

…我的好友

×一FeraiDashboerd禮物

?22202m

C

3Frefsxx香看我的朋理書(shū)用一我是誰(shuí)他的東西社區(qū)平臺(tái)他的活動(dòng)2/16/2011

Ed

Chang

20開(kāi)

放

社

區(qū)

平

臺(tái)我

的

朋

友e

o:c-mmgrouppnetoa2?2210*

風(fēng)*

4015程片②關(guān)置來(lái)花要開(kāi)復(fù)的禮物6ktdryeeHstryEoomartsSe開(kāi)復(fù)四實(shí)來(lái)來(lái)地開(kāi)復(fù)的要集積片6t

6dt·~iGoooe

cn-Roos(a)-edcrgo

…1oosHnp:/hba,trraanuba/OFtVeN-151308031N3*1c8oK

**e-s

#州2讀的禮物tCst.的力收到1出

0LecMePFLm

3!要

的ula×c-他的主開(kāi)復(fù)收到的禮物曰

?驗(yàn)明

進(jìn)出&o

天準(zhǔn)書(shū)形·

我的繳東細(xì)回

C貿(mào)意startO2me-

開(kāi)案的民物孔品廟禮物’*nmAPo送出0menesDone21Social

GraphU

i

t

r

NPi294NANomViee103

KinuteBrowsePicturesUnie…-ViVseei

127

Im31Time

Ter

V

23

Nne=mo

intense

largestcirclerepre

entJoinorVisit

GroupsUni

iVt

st

s

i

li

nlllonrime

Per

Visit

!3s

MinutesBrowseMarketPlaceUn*a

sYtis=te2e

il,

n

*enAddAFriendUi

i

?s*y

6

Mi2l2

t

e*CAC

mp

et

comlargestaudienceis*tmoAn

inMtu…io3sineoadeSewevtember(BOr

soer

PFroiefnil

s)

Unieus

Visitem:21

Million

ri

V=*

.Vi-i26e2

a:iSM

Mon

inutesloin

or

Browse

NetworksRead

Discussion

Boards

a-uP

iVstia

s

%u2ttnSearch

for

Members

and

GroyPs

VuiP-…

ri…

iti

s

s

*0*4e7$mitl

MittesomniTiUueellioimniTU

o

e

t

ot5em64in

rl

ta

i

tnhsUniawe

Visitors:

14envime

Pr

Visaa

e

4=70

nutsneottTimUn

zce

e

ek:

t

tyStBalr

dbus

h2

d0e2represent

usage

intesityonrweouanesvncEd

Chang

23Darkeshade2/16/2011T

sri

:

o

M1

$i

suteominnt1:*ssimUWhat

UsersWant?·People

care

about

other

people一careabout

peoplethey

know·

一

connect

to

people

they

do

not

know一

about

who

other

people

are一aboutwhatotherpeople

are

doingDiscover

interesting

information一

basedonother

people2/16/2011

Ed

Chang

24InformationOverflow

ChallengeT

u

n

n

,too

many

choices

of·DesiringaSocialNetworkRecommendation

System2/16/2011

Ed

Chang

25appseopleaysammforoo··“/

ds

i

e

me

to

manageorkfulltwanreliaheenlinneonmysooRecommendationSystemu

n

mmendationoncoRdaForumommeyenitRComFrien···Application

Suggestion

·Ads

Matching2/16/2011

Ed

Chang

26Organizing

the

World's

Information,Socially·

社

區(qū)平

臺(tái)(Social

Platform)·云

運(yùn)

算(Cloud

Computing)·結(jié)論與前瞻(Concluding

Remarks)2/16/2011Ed

Chang27(3

)

算(4

的云計(jì)算空強(qiáng)無(wú)限無(wú)限··)是你的的云計(jì)就備在后設(shè)不錄何所登任無(wú)··(1)數(shù)據(jù)在云端·不怕丟失·不必備份(2

)

端升下級(jí)載在云動(dòng)必件不自軟··業(yè)界趨勢(shì)：云計(jì)算時(shí)代的到來(lái)無(wú)限速度互聯(lián)網(wǎng)搜索：

云計(jì)算的例子2.分布式預(yù)處理數(shù)據(jù)以便為搜索提供服務(wù)：

Gcogle

Infrastructure(thousands

sesdi

e

for

mass

data一

Google

FileSystem一ngucproMa—taVeteateTatH* *

RPmutm-CooeA

Clto*eC.OteSeeCiePit*1邊

a7004#Cemmm

-hm1

rmCeat

conputs?-

Google

Cloud

Computing

inraetrguesy

BAEd

CSsnouhSm29

CsOaltm

Jn

msE+eaooglIn

limodity

servers

arcund

theworldofco1.用戶輸入查詢關(guān)鍵字3.返回搜索結(jié)果2/16/2011netEelermCemuoudtse2145541335245341352141554254331521312345133352115241355125Collaborative

FilteringGiven

a

matrix

that“encodes”data2/16/2011

Ed

Chang

30214554?133524?53?413521?455425?2

4335213123451333?52?1152?4435451245?Given

a

matrix

that“encodes

”dataManyapplications·User-Community·User-User·Ads

-User·Ads-Community·etc.(collaborativefiltering):Ed

Chang

31Communities2/16/2011UsersCollaborative

Filtering(CF)[Breese,Heckerman

and

Kadie

1998]·Memory-based—

h

il

fi

sm

i

r

,

)·Model-based—Build

a

model

of

relationship

between

subject

matters一Make

predictions

basedonthe

constructedmodelcstreolehprofineiglasts,simieareenrrilasemuwsiilar”samssr,aumrsisetugnBouiveDifferent

similarity

measures

yield

different

techniques一

ons

based

on

the

preferences

of

theseersictiueimilarake

p“sM2/16/2011

Ed

Chang

32Memory-Based

Model[Goldbertetal.1992;Resniket

al.1994;Konstant

et

al.1997]·

Pros一

Simplicity,avoid

model-building

stage·

Cons—MemoryandTimeconsuming,uses

the

entiredatabaseeverytimetomake

aprediction一

Cannotmake

prediction

ifthe

user

has

no

items

incommonwithother

users2/16/2011

Ed

Chang

33Model-Based

Model[Breese

et

al.1998;Hoffman

1999;Blei

et

al.2004]Pros一

it

t

is

much

smaller

than

the

一

nti

ery

the

model

instead

ofCons一

Model-buildingtakestimedatasetdiction,quererpeertheFastatasemodelllabiactuScala2/16/2011

Ed

Chang

34Algorithm

Selection

Criteria

S

al

i

commendation·CloudComputing!ngReainmeTr-ticalableNear-re·

e

t

it

r

i

gs

r

irablecityDesasciantianhTwaldealmenCanIncr2/16/2011

Ed

Chang

35Model-based

PriorWorkLatent

Semantic

Analysis

(LSA)·

ProbabilisticLSA(PLSA)··Latent

Dirichlet

Allocation(LDA)2/16/2011

Ed

Chang

36·

Maphigh-dimensional

count

vectors

tolowerdimensional

representation

called

latent

semantic

space·BySVD

decomposition:A=UEVTDocs

Word×D

SWxD

WxTA=Word-document

co-occurrence

matrixU;

=How

likely

word

i

belongs

to

topic

jji

=How

significant

topic

j

isVi1=How

likely

topic

i

belongs

to

docjLatent

Semantic

Analysis(LSA)[Deerwester

et

al.1990]2/16/2011

Ed

Chang

37

Latent

Semantic

Analysis(cont.)·LSAkeepsk-largestsingular

values一

Low-rankapproximationtothe

original

matrix一

Savespace,de-noisifiedandreducesparsityOCS·Make

ecommendations

usingA—Word-word

similarity:A

AT-Doc-doc

similarity:?T

A—Word-doc

relationship:AWxKWxrDATopKxD2/16/2011ChangWordsEd38Probabilistic

Latent

Semantic

Analysis(PLSA)[Hoffman

1999;Hoffman20041Document

is

viewed

as

a

bag

ofwords··

|

,

licit

meaningiEMexpviwrsPlingd),deea,Plw)ityodel-P(robaMP··A

latent

semantic

layer

isconstructed

inbetweendocuments

and

words

·P(w,d)=P(d)P(w|d)=P(d)EzP(w|z)P(z|d)algorithm2/16/2011ChangEd39·

LDA[Blei

et

al.2003]一

Provideacompletegenerativemodelwith

Dirichlet

prior

·

AT

[Griffiths

&Steyvers

2004]一

Includeauthorshipinformation一

Document

iscategorizedbyauthors

andtopics·

ART[McCallum2004]一

Includeemailrecipientas

additional

information一

Email

is

categorized

by

author,recipients

andtopics2/16/2011

Ed

Chang

40PLSAextensions·

PHITS[Cohn

&Chang

2000]·

e

t

u

[Cohn

&Hoffmann

2001]一

Model

contents(words)and

inter-connectivity

of

documentsHITSencePrrSA

andco-occLnPiombinationofocument-citaclA

li一CombinationalCollaborativeFiltering(CCF)·Fusemultiple

information—Alleviate

the

information

sparsity

problem·Hybridtrainingscheme—Gibbs

sampling

as

initializations

for

EM·Parallelization—Achieve

linear

speedup

with

the

numberof

machines2/16/2011

Ed

Chang

41algorithm·

Givenacollectionofco-occurrencedata-Community:C

={C?

,C?

,…,C}-User:U={u?

,u?

,…,um}一Description:D={d?

,d?

,…,dv}-Latentaspect:Z={z?

,z?

,…,zk}·Models—Baseline

models·Community-User(C-U)model·Community-Description(C-D)model-CCF:CombinationalCollaborativeFiltering·

Combines

both

baseline

modelsNotations2/16/2011ChangEd42·Communityis

viewed

as

a

bag

ofwords

·canddarerendered

conditionally·Gi

ent

pr

rodu

zword

d1.A

community

cischosenuniformly

2.A

topic

zisselectedfromP(z|c)3.Awordd

isgenerated

from

P(d|z)Chang

43hgcnaiecrtonss,firativpeneeen·Communityisviewed

as

a

bag

of

usersc

and

u

are

rendered

conditionallyindependentbyintroducingz■(

Generative

process,for

each

user

u

1.Acommunitycischosenuniformly

2.A

topic

zisselected

fromP(z|c)3.Auser

u

is

generatedfrom

P(u|z)2/16/2011

EdModelsCommunity-Description(C-D)model

BaselineCommunity-User(C-U)model

-Pros1.Cluster

communities

based

oncommunity

content(description

words)-Cons1.No

personalized

recommendation2.Donot

considerthe

overlapped

usersbetween

communitiesChang

441.C-U

matrix

information2.Cannot

take

similarity2/16/2011is

sparse,may

sufferfromsparsity

problemadvantage

of

contentbetween

communitiesEdModels(cont.)Community-Description(C-D)model

BaselineCommunity-User(C-U)model

-Pros1.Personalized*Conssuggestioncommunity·CCFcombines

both

baseline

models*A

community

isviewed

as-abag

of

users

AND

a

bag

ofwords*By

adding

C-U,CCF

can

performpersonalizedrecommendationwhichC-Dalone

cannot·By

adding

C-D,CCF

can

perform

betterpersonalizedrecommendationthan

C-Ualonewhich

may

sufferfrom

sparsity·Things

CCF

can

do

that

C-U

and

C-Dcannot-P(d)u),relate

user

to

word-Useful

for

user

targeting

adsCombinational

Filtering(CCF)model

C

P(c)—P(zlc)ZP(ulz)

P(dlz)U

dModelCCF2/16/2011CollaborativeChangEd45Algorithm

Requirements

S

al

i

commendationngReainmeTr-ticalableNear-reIncrementalTraining

is

Desirable2/16/2011

Ed

Chang

46ParallelizingCCFDetailsomitted2/16/2011

Ed

Chang

47(3

)

算(4

的云計(jì)算空強(qiáng)無(wú)限無(wú)限··)是你的的云計(jì)就備在后設(shè)不錄何所登任無(wú)··(1)數(shù)據(jù)在云端·不怕丟失·不必備份(2

)

端升下級(jí)載在云動(dòng)必件不自軟··業(yè)界趨勢(shì)：云計(jì)算時(shí)代的到來(lái)無(wú)限速度ExperimentsonOrkut

Dataset·Data

description-Collected

on一

Two

types

ofJuly

26,2007data

were

extracted·Community-user,community-description

一312,385users—109,987communities·

—191,034

unique

Englishwords·Speedup·

Community

recommendation·U

m

i

milarity/clusteringtysiartylisimimunserCo2/16/2011Ed

Chang49Community

Recommendation·

Evaluation

Method一

No

ground-truth,no

user

clicks

available—Leave-one-out:randomly

delete

one

community

foreach

user一

Whether

the

deleted

community

can

be

recovered·

Evaluation

metric—Precisionand

Recall2/16/2011

Ed

Chang

50Lengthoftherecommendation

listPecertageObservations:

CCFoutperforms

C-U

e

cboe

u

Citi

C

Uu

nhascaserttermmeehr,todmjoinThe口

For

top20,precision/recall

of

CCFare

twice

higher

than

those

of

C-UNumber

of

communities

a

user

has

joinedEpredict2/16/2011Ed

Chang51·

The

Orkut

dataset

enjoys

a

linear

speedup

when

the

number

of2/16/2011

Ed

Chang

52machines

is

up

to

100

Reduces

the

training

time

from

one

day

to

less

than

14

minutes··RuntimeSpeedupMachinesTime(aee.)Specdup100.23310204,32621.3502.28040.510O1,01491.1200706116But,what

makes

the

speedup

slow

down

after

100

machines?Number

of

mnchinesSpeedup200RuntimeSpeedup(cont.)·Trainingtimeconsistsoftwo

parts:一

Computationtime(Comp)一

Communicationtime(Comm)sdoedupNumberofmachinesNumbarofmachines2/16/2011

Ed

Chang

53CCFSummary·CombinationalCollaborative

Filtering—Fuse

bags

ofwordsand

bags

of

usersinformation—Hybridtrainingprovides

better

adliitionsfor

EM

ratherthan

random一

Parallelizeto

handle

large-scaledatasetsngzaetiseini2/16/2011

Ed

Chang

54China'sContributionson/to

CloudComputing

Parallel

CCF Parallel

SVMs(Kernel

Machines)·

ParallelSpectral

Clustering·

Parallel

Expectation

Maximization

·ParallelAssociation

Mining·

Parallel

LDA2/16/2011

Ed

Chang

55Parallel

SVDSpeeding

up

SVMs

[NIPS

2007]·Approximate

MatrixFactorization·

Parallelization

Open

source

@/p/psvm·A

task

that

takes

7

dayson

1machinetakes

1

hourson500

machines350+downloads

since

December

072/16/2011

Ed

Chang

56≈XIncompleteCholesky

Factorization(ICF)p<<n→Conserve

Storage2/16/2011

Ed

Chang