1 Crawling the Web Discovery and Maintenance of Large-Scale Web Data Junghoo Cho Stanford University.

1

Crawling the WebCrawling the Web

Discovery and Maintenance of Discovery and Maintenance of

Large-Scale Web DataLarge-Scale Web Data

Junghoo ChoJunghoo Cho

Stanford UniversityStanford University

2

What is a Crawler?What is a Crawler?

web

init

get next url

get page

extract urls

initial urls

to visit urls

visited urls

web pages

3

ApplicationsApplications

Internet Search EnginesInternet Search Engines– Google, AltaVistaGoogle, AltaVista

Comparison Shopping ServicesComparison Shopping Services– My Simon, BizRateMy Simon, BizRate

Data miningData mining– Stanford Web Base, IBM Web FountainStanford Web Base, IBM Web Fountain

4

WebBase CrawlerWebBase Crawler

Web Base ProjectWeb Base Project BackRub Crawler, PageRankBackRub Crawler, PageRank GoogleGoogle New Web Base CrawlerNew Web Base Crawler

– 20,000 lines in C/C++20,000 lines in C/C++– 130M pages collected130M pages collected

5

Crawling Issues (1)Crawling Issues (1)

Load at visited web sitesLoad at visited web sites– Space out requests to a siteSpace out requests to a site– Limit number of requests to a site per dayLimit number of requests to a site per day– Limit depth of crawlLimit depth of crawl

6

Crawling Issues (2)Crawling Issues (2)

Load at crawlerLoad at crawler– ParallelizeParallelize

init

get next url

get page

extract urls

initial urls

to visit urls

visited urls

web pages

init

get next url

get page

extract urls

?

7

Crawling Issues (3)Crawling Issues (3)

Scope of crawlScope of crawl– Not enough space for “all” pagesNot enough space for “all” pages– Not enough time to visit “all” pagesNot enough time to visit “all” pages

Solution: Visit “important” pages

visitedpages

Intel

Intel

8

Crawling Issues (4)Crawling Issues (4)

ReplicationReplication– Pages mirrored at multiple locationsPages mirrored at multiple locations

9

Crawling Issues (5)Crawling Issues (5)

Incremental crawlingIncremental crawling– How do we avoid crawling from scratch?How do we avoid crawling from scratch?– How do we keep pages “fresh”?How do we keep pages “fresh”?

10

Summary of My ResearchSummary of My Research

Load on sites [PAWS00]Load on sites [PAWS00] Parallel crawler [Tech Report 01]Parallel crawler [Tech Report 01] Page selection [WWW7]Page selection [WWW7] Replicated page detection [SIGMOD00]Replicated page detection [SIGMOD00] Page freshness [SIGMOD00]Page freshness [SIGMOD00] Crawler architecture [VLDB00]Crawler architecture [VLDB00]

11

Outline of This TalkOutline of This Talk

How can we maintain pages fresh?How can we maintain pages fresh? How does the Web change?How does the Web change? What do we mean by “fresh” pages?What do we mean by “fresh” pages? How should we refresh pages?How should we refresh pages?

12

Web Evolution ExperimentWeb Evolution Experiment

How often does a Web page change?How often does a Web page change? How long does a page stay on the Web?How long does a page stay on the Web? How long does it take for 50% of the Web How long does it take for 50% of the Web

to change?to change? How do we model Web changes?How do we model Web changes?

13

Experimental SetupExperimental Setup

February 17 to June 24, 1999February 17 to June 24, 1999 270 sites visited (with permission)270 sites visited (with permission)

– identified 400 sites with highest “PageRank”identified 400 sites with highest “PageRank”– contacted administratorscontacted administrators

720,000 pages collected720,000 pages collected– 3,000 pages from each site daily3,000 pages from each site daily– start at root, visit breadth first (get new & old pages)start at root, visit breadth first (get new & old pages)– ran only 9pm - 6am, 10 seconds between site requestsran only 9pm - 6am, 10 seconds between site requests

14

Average Change IntervalAverage Change Intervalfr

actio

n of

pag

es

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

1day 1day- 1week

1week-1month

1month-4months

4months

average change interval

15

Change Interval – By DomainChange Interval – By Domainfr

actio

n of

pag

es

0

0.1

0.2

0.3

0.4

0.5

0.6

1day 1day- 1week

1week-1month

1month-4months

4months

com

netorg

edu

gov

average change interval

16

Modeling Web EvolutionModeling Web Evolution

Poisson process with rate Poisson process with rate T is time to next eventT is time to next event ffTT ((tt) = ) = ee--

tt ( (tt > 0) > 0)

17

Change Interval of PagesChange Interval of Pagesfor pages thatchange every

10 days on average

interval in days

frac

tion

of c

hang

esw

ith g

iven

inte

rval

Poisson model

18

Change MetricsChange Metrics

FreshnessFreshness– Freshness of element Freshness of element eeii at time at time tt is is

F F ( ( eeii ; ; tt ) = 1 if ) = 1 if eeii is up-to-date at time is up-to-date at time tt 0 otherwise 0 otherwise

eiei

......

web databaseFreshness of the database S at time t is

F( S ; t ) = F( ei ; t )

(Assume “equal importance” of pages)

N

1 N

i=1

19

Change MetricsChange Metrics

AgeAge– Age of element Age of element eeii at time at time tt is is

A A( ( eeii ; ; tt ) = 0 if ) = 0 if eeii is up-to-date at time is up-to-date at time tt tt - (modification - (modification eei i time) otherwisetime) otherwise

eiei

......

web databaseAge of the database S at time t is

A( S ; t ) = A( ei ; t )

(Assume “equal importance” of pages)

N

1 N

i=1

20

Change MetricsChange Metrics

F(ei)

A(ei)

0

0

1

time

time

update refresh

Time averages:

0

1( ) lim ( ; )

t

i itF e F e t dt

t

0

1( ) lim ( ; )

t

tF S F S t dt

t

21

Refresh OrderRefresh Order

Fixed orderFixed order– Explicit list of URLs to visitExplicit list of URLs to visit

Random orderRandom order– Start from seed URLs & follow linksStart from seed URLs & follow links

Purely randomPurely random– Refresh pages on demand, Refresh pages on demand, as requested by useras requested by user

eiei

......

webdatabase

22

Freshness vs. Revisit FrequencyFreshness vs. Revisit Frequency

r = / f = average change frequency / average visit frequency

23

Age vs. Revisit FrequencyAge vs. Revisit Frequency

r = / f = average change frequency / average visit frequency

= Age / time to refresh all N elements

24

Trick QuestionTrick Question

Two page databaseTwo page database e1 changes dailychanges daily e2 changes once a weekchanges once a week Can visit one page per weekCan visit one page per week How should we visit pages?How should we visit pages?

– e1 e2 e1 e2 e1 e2 e1 e2... ... [uniform] [uniform]

– e1 e1 e1 e1 e1 e1 e1 e2 e1 e1 … … [proportional][proportional]

– e1 e1 e1 e1 e1 e1 ... ...

– e2 e2 e2 e2 e2 e2 ... ...

– ??

e1

e2

e1

e2

webdatabase

25

Proportional Often Not Good!Proportional Often Not Good!

Visit fast changing Visit fast changing e1

get 1/2 day of freshnessget 1/2 day of freshness

Visit slow changing Visit slow changing e2

get 1/2 week of freshnessget 1/2 week of freshness

Visiting Visiting e2 is a better deal!is a better deal!

26

Optimal Refresh FrequencyOptimal Refresh Frequency

ProblemProblem

Given and Given and f ,f ,

findfind

that maximizethat maximize

1 2, , ..., N

1 21

, ,... , /N

N

ii

f f f f f N

1

1( ) ( )

N

ii

F S F eN

27

SolutionSolution

Compute Compute

Lagrange multiplier methodLagrange multiplier method

AllAll

( )iF e

1 /

0

1

0

1 1( ) ( ; )lim

/

ii ff

n

i in j i i

j eF e f F e t dt

n f f

/ /( , )( ) 1i i i if fi i

i i i i

F fF S e e

f f f

/ /1

( , ) pairs should satisfy f f

i ie e

ff

28

Optimal Refresh FrequencyOptimal Refresh Frequency

• Shape of curve is the same in all cases• Holds for any change frequency distribution

29

Optimal Refresh for AgeOptimal Refresh for Age

• Shape of curve is the same in all cases• Holds for any change frequency distribution

30

Comparing PoliciesComparing Policies

Freshness AgeProportional 0.12 400 days

Uniform 0.57 5.6 daysOptimal 0.62 4.3 days

Based on Statistics from experimentand revisit frequency of every month

31

Topics to FollowTopics to Follow

Weighted FreshnessWeighted Freshness Non-Poisson ModelNon-Poisson Model Change Frequency EstimationChange Frequency Estimation

32

Not Every Page is Equal!Not Every Page is Equal!

1 211

( ) ( ) (2 )3

F S F e F e

In general,1 1

( ) ( )N N

i i ii i

F S w F e w

e1

e2 Accessed by users 20 times/day

Accessed by users 10 times/day

Some pages are “more important”Some pages are “more important”

33

Weighted FreshnessWeighted Freshness

w = 1

w = 2

f

34

Non-Poisson ModelNon-Poisson Model

interval in days

frac

tion

of c

hang

esw

ith g

iven

inte

rval

Poisson model

Heavy-tail distribution 2(1 )t

35

Optimal Revisit FrequencyOptimal Revisit Frequencyfor Heavy-Tail Distributionfor Heavy-Tail Distribution f

36

Principle of Diminishing ReturnPrinciple of Diminishing Return

TT: time to next change: time to next change : continuous, differentiable: continuous, differentiable Every page changesEvery page changes

Definition of change rate Definition of change rate

Pr( | )T t

Pr( 0 | ) 1 and Pr( | ) 0limt

T T t

Pr( | ) Pr( | )T t k T kt

0lim f

37

Change Frequency EstimationChange Frequency Estimation

How to estimate change frequency?How to estimate change frequency?– Naïve Estimator: Naïve Estimator: XX//TT

– XX: number of detected changes: number of detected changes

– TT: monitoring period: monitoring period

– 2 changes in 10 days: 0.2 times/day2 changes in 10 days: 0.2 times/day

Change detected1 day

Page visitedPage changed

Incomplete change historyIncomplete change history

38

Improved EstimatorImproved Estimator

Based on the Poisson modelBased on the Poisson model

– XX: number of detected changes: number of detected changes– NN: number of accesses: number of accesses– f f : access frequency: access frequency

2log

1

Nf

N X

3 changes in 10 days: 0.36 times/day Accounts for “missed” changes

39

Improved EstimatorImproved Estimator

BiasBias

EfficiencyEfficiency

ConsistencyConsistency

0

2ˆ[ ] log 1

1

N iN ir r

i

NNE r e e

iN i

0

22

ˆ[ ] log 11

N iN ir r

i

NNVar r e e

iN i

0

22

log 11

N iN ir r

i

NNe e

iN i

ˆ ˆE[ ] Var[ ] 0lim limn n

r r r

40

Improvement Significant?Improvement Significant?

Application to a Web crawlerApplication to a Web crawler– Visit pages once every week for 5 weeksVisit pages once every week for 5 weeks– Estimate change frequency Estimate change frequency – Adjust revisit frequency based on the estimateAdjust revisit frequency based on the estimate

» Uniform: do not adjustUniform: do not adjust

» Naïve: based on the naïve estimatorNaïve: based on the naïve estimator

» Ours: based on our improved estimatorOurs: based on our improved estimator

41

Improvement from Our EstimatorImprovement from Our Estimator

Detected changesDetected changes Ratio to uniformRatio to uniform

UniformUniform 2,147,5892,147,589 100%100%

NaïveNaïve 4,145,5824,145,582 193%193%

OursOurs 4,892,1164,892,116 228%228%

(9,200,000 visits in total)

42

Other EstimatorsOther Estimators

Irregular access intervalIrregular access interval Last-modified dateLast-modified date CategorizationCategorization

43

SummarySummary

Web evolution experimentWeb evolution experiment Change metricChange metric Refresh policyRefresh policy Frequency estimatorFrequency estimator

44

ContributionContribution

Freshness [SIGMOD00]Freshness [SIGMOD00] Page selection [WWW7]Page selection [WWW7] Replicated page detection [SIGMOD00]Replicated page detection [SIGMOD00] Load on sites [PAWS00]Load on sites [PAWS00] Parallel crawler [Tech Report 01]Parallel crawler [Tech Report 01] Crawler architecture [VLDB00]Crawler architecture [VLDB00]

45

The EndThe End

Thank you for your attentionThank you for your attention For more information visitFor more information visit

http://www-db.stanford.edu/~cho/http://www-db.stanford.edu/~cho/