Evaluating Sliding and Sticky Target Policies by Measuring Temporal Drift in Acyclic Walks Through a Web Archive�

EVALUATING SLIDING AND STICKY TARGET POLICIES BY MEASURING TEMPORAL DRIFT IN ACYCLIC WALKS THROUGH A WEB ARCHIVE

SCOTT G. AINSWORTH

MICHAEL L. NELSON

OLD DOMINION UNIVERSITY

COMPUTER SCIENCE

JCDL 2013

JULY 23-25, 2013

INDIANAPOLIS, INDIANA USA

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013

Scott G. Ainsworth • Michael L. Nelson

2

A FABLE FROM WAYBACK

7/23/13

A long, long time ago…

ODU Computer Scienceupdated its web site…

What did it look like?

May 2005...

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


3


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


4


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


5


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


6


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


7

WHAT JUST HAPPENED?WHAT WE EXPECTED2005-05-14 @ 01:36:08

WHAT WE GOT2005-03-31 @ 09:16:10

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


8

SLIDING TARGET

7/23/13

2005-05-14 01:36:08

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


9

SLIDING TARGET

7/23/13

2005-04-2200:17:52

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


10

SLIDING TARGET

7/23/13

2005-03-3109:16:10

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


11

STICKY TARGET

What if the target is held steady?

(Enabled by Memento API)

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


12

MEMENTO HTTP EXTENSION*Adds ability to request a particular date-timeEnables Sticky Target

Request

Response

7/23/13

GET <timegate>/http://www.cs.odu.edu/ HTTP/1.1…Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT…

HTTP/1.1 200 OK…Memento-Datetime: Sat, 14 May 2005 01:36:08 GMT…

*https://datatracker.ietf.org/doc/draft-vandesompel-memento/

https://datatracker.ietf.org/doc/draft-vandesompel-memento/

https://datatracker.ietf.org/doc/draft-vandesompel-memento/

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


13

2005-05-142005-05-1401:36:08

STICKY TARGET

7/23/13

Mem

ento

Fo

x E

xten

sio

n

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


14

STICKY TARGET

7/23/13

2005-04-2200:17:52

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


15

STICKY TARGET

7/23/13

2005-05-1401:36:08

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


16

DRIFT COMPARISON

PageSliding Sticky

Datetime Drift Datetime Drift

CS Home2005-05-14

01:36:08– 2005-05-14

01:36:08–

Science Home

2005-04-2200:17:52

22.1 days 2005-04-2200:17:52

22.1 days

CS Home2005-03-31

09:16:1043.7 days(+21.6 days)

2005-05-1401:36:08

–

Mean 32.9 days 11.0 days

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


17

QUESTIONSHow much temporal drift is there with the two policies?

Does the sticky policy reduce drift as expected?

If so, by how much?

How do • Choice (number of links)• Domains visited• Walk length

Influence drift?

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


18

CONTENTS Motivation

Related work

Measuring Drift

Results

Future work

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


19

RELATED WORKControl Crawl Data Quality, Future collections

• Spaniol et al. – crawling strategy • Denev et al. – change rates by MIME type and

depth• Ben Saad et al. – metadata from crawl used to

select best results from archive

Our Focus: Existing Data Quality• Existing collections• Datetime selection policies

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


20

CONTENTS Motivation

Related work

Measuring drift

Results

Future work & conclusions

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


21

DEFINITIONS

Walk Length Number of successful steps (HTTP 200 response)

Unique Domains

Number of unique domains (jcdl.org, amazon.com, etc.)

Choice Number of unique links (calculated per page)

Drift | target-datetime1 – Memento-Datetimei |

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


22

PROCESS BY EXAMPLESelect a URI

• Random selection of 1 out of 4,000

4000 Sample URIs – same as JCDL 2011 paper• DMOZ – a reference• Search Engines – best random sampling• Bitly – does shortening have an impact?• Delicious – does popularity have an impact?

“How Much of the Web Is Archived?”

http://arxiv.org/abs/1212.6177

7/23/13

http://arxiv.org/abs/1212.6177

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


23

PROCESS BY EXAMPLEFirst, select a URI

• Random selection of 1 out of 4,000

Second, download timemap

7/23/13

<http://api.wayback.archive.org/memento/20050507093740/http://www.cs.odu.edu/>; rel="memento"; datetime="Sat, 07 May 2005 09:37:40 GMT", <http://api.wayback.archive.org/memento/20050514013608/http://www.cs.odu.edu/>; rel="memento"; datetime="Sat, 14 May 2005 01:36:08 GMT", <http://api.wayback.archive.org/memento/20050515002903/http://www.cs.odu.edu/>; rel="memento"; datetime="Sun, 15 May 2005 00:29:03 GMT",

<http://api.wayback.archive.org/memento/20050514013608/http://www.cs.odu.edu/>; rel="memento"; datetime="Sat, 14 May 2005 01:36:08 GMT",

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


24

PROCESS BY EXAMPLENext, download both mementos

Wayback Machine Memento API

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


25

PROCESS BY EXAMPLENext, download both mementos

And Find common links


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


26

STATUS SO FAR

Successful Steps 1

Unique Domains 1

Choice 48

Mean Drift (days) 0.0 WB 0.0 API

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


27

PROCESS BY EXAMPLEFind common links

and select one for the next step


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


28

PROCESS BY EXAMPLEThe timemap downloaded, the best datetimes are selected, and the memento downloaded…


7/23/13

Successful Steps 1 + 1 = 2

Unique Domains 1 + 0 = 1

Choice 48 + 36 = 84


Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


29

PROCESS BY EXAMPLEAgain for http://www.odu.edu


7/23/13



Choice 84 + 33 = 117


http://www.odu.edu/

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


30

HTTP Response:

• 302 Redirect• Location header

PROCESS BY EXAMPLEAnd for http://www.odusports.com

Redirected at acquisition time


7/23/13

http://www.odusports.com/

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


31

PROCESS BY EXAMPLEAnd for http://odusports.collegesports.com


7/23/13



Choice 117 + 77 = 194


http://odusports.collegesports.com/

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


32

PROCESS BY EXAMPLEAnd for http://www.vtext.com


7/23/13



Choice 194 + 14 = 208


http://www.vtext.com/

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


33

PROCESS BY EXAMPLEAnd 404 stops the walk


7/23/13

HTTP Response:

• 404 Not Found



Choice 194 + 14 = 208


Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


34

STOP CAUSESFirst Step Subsequent Steps

Stop Cause Count Percent Count Percent

Timemaps

HTTP 403 74 1.7% 4,803 9.1%

HTTP 404 1,327 30.1% 15,850 29.0%

HTTP 503 0 0.0% 43 0.1%

Other 2 0.0% 180 0.3%

Mementos

HTTP 403 52 1.2% 476 0.9%

HTTP 404 215 4.9% 3,633 6.8%

HTTP 503 1,957 44.4% 10,535 19.9%

Download failed 154 3.5% 589 1.1%

Not HTML 514 11.7% 2,856 5.4%

No Common Links 0 0.0% 12,957 24.4%

Other 117 2.7% 1,128 2.1%

Totals 4,412 53,050

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


35

CONTENTS Motivation

Related work

Measuring drift

Results


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


36

WALKS AND STEPS

Status Total

Walks Attempted 200,000

Unique Walks 53,100

Successful Walks 48,685

Pct. Successful 91.7%

Steps 240,439

Successful Steps 187,371

w/drift > 1yr 6,701

w/drift > 5yrs 111

Successful Steps/Walk 3.8

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


37

WALK LENGTHS

7/23/13

Walk Length

Occ

urre

nces

(lo

g sc

ale)

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


38

MEDIAN DRIFT BY STEP

● Sliding● Sticky

Med

ian

Drif

t (m

onth

s)

7/23/13

Step Number

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


39

DRIFT BY STEP

SLIDING POLICY STICKY POLICY

Drif

t (y

ears

)

Step Number Step Number

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


40

DRIFT BY CHOICE

7/23/13

Choice

Mea

n D

rift

(mon

ths)


Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


41

DRIFT BY DOMAINS

7/23/13

Domain Count

Mea

n D

rift

(mon

ths)


Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


42

CONTENTS Motivation

Related work

Measuring drift

Results


7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


43

FUTURE WORKIntegrate real-world walk patterns

• AlNoamany et al. – Internet Archive logs• Domains users avoid – link farms, etc.• Domain clusters• Self referencing domains – 101celebrities.com

Check other archives• Other archives now have Memento API

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


44

CONCLUSIONS

30 days less drift using Sticky policy.

Sticky policy controls drift;Sliding policy does not.

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


45

BACKUP

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


46

WALK LENGTHSWalk Length DMOZ S.Eng. Delicious Bitly Total

1 5,355 1,239 7,139 1,289 15,076

2 3,571 924 4,857 817 10,169

3 1,891 598 3,311 623 6,423

4 1,212 381 2,228 415 4,236

5 791 315 1,588 314 3,008

6 583 232 1,168 259 2,242

7 417 178 877 186 1,658

8 258 153 651 136 1,198

9 187 111 498 108 904

10 144 79 337 79 679

…

20 14 10 36 9 76

…

41-45 6 2 14 2 24

46-50 6 3 6 1 16

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


47

MEAN DRIFT BY STEP

7/23/13

Step Number

Mea

n D

rift

(mon

ths)


● μ ○ σ

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


48

SLIDING TARGET⟹ GET …/20050514013608/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET …/20050514013608/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 302 FOUND

Location: …/20050522001752/http://sci.odu.edu/⟹ GET …/20050522001752/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET …/20050522001752/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 302 FOUND

Location: …/20050331091610/http://www.cs.odu.edu/⟹ GET …/20050331091610/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


49

SLIDING TARGET⟹ GET …/20050514013608/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET …/20050514013608/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 302 FOUND

Location: …/20050522001752/http://sci.odu.edu/⟹ GET …/20050522001752/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET …/20050522001752/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 302 FOUND


7/23/13

22 Days

44 Days

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


50

STICKY TARGET⟹ GET <timegate>/http://www.cs.odu.edu/ HTTP/1.1

Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT⟸ HTTP/1.1 302 FOUND


⟹ GET <timegate>/http://sci.odu.edu/ HTTP/1.1Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT

⟸ HTTP/1.1 302 FOUNDLocation: …/20050522001752/http://sci.odu.edu/

⟹ GET …/20050522001752/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET <timegate>/http://www.cs.odu.edu/ HTTP/1.1Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT

⟸ HTTP/1.1 302 FOUNDLocation: …/20050514013608/http://www.cs.odu.edu/

⟹ GET …/20050514013608/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


51

STICKY TARGET (MEMENTO)⟹ GET <timegate>/http://www.cs.odu.edu/ HTTP/1.1

Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT⟸ HTTP/1.1 302 FOUND


⟹ GET <timegate>/http://sci.odu.edu/ HTTP/1.1Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT

⟸ HTTP/1.1 302 FOUNDLocation: …/20050522001752/http://sci.odu.edu/

⟹ GET …/20050522001752/http://sci.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

⟹ GET <timegate>/http://www.cs.odu.edu/ HTTP/1.1Accept-Datetime: Sat, 10 May 2005 11:21:00 GMT

⟸ HTTP/1.1 302 FOUNDLocation: …/20050514013608/http://www.cs.odu.edu/

⟹ GET …/20050514013608/http://www.cs.odu.edu/ HTTP/1.1⟸ HTTP/1.1 200 OKAY

7/23/13

22 Days

0 Days

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


52

TWO BROWSING POLICIES

SLIDING TARGET

Target• Resource datetime

Drift types• Memento drift• Target drift

STICKY TARGET

Target• Original datetime

Drift type• Only memento drift

7/23/13

Join

t C

onfe

renc

e o

n D

igita

l Lib

rarie

s (J

CD

L) 2

013


53

TWO TYPES OF DRIFTTarget Drift

• Drift introduced by changing the target datetime• | received-datetime – original-datetime |

Memento Drift• Drift introduced by not having the exact datetime

requested available.• | received-datetime – requested-datetime |

7/23/13

Evaluating Sliding and Sticky Target Policies by Measuring Temporal Drift in Acyclic Walks Through a Web Archive�

Technology

ainsworth michael

web archive scott

sticky target policies

sticky target request

sticky policy

days cs home

measuring temporal drift

ok mementodatetime