Web search Web data management and distribution Serge Abiteboul Ioana Manolescu Philippe Rigaux Marie-Christine Rousset Pierre Senellart

(1)

Web search

Web data management and distribution

Serge Abiteboul Ioana Manolescu Philippe Rigaux Marie-Christine Rousset Pierre Senellart

Web Data Management and Distribution http://webdam.inria.fr/textbook

September 18, 2013

(2)

Outline

1 Web crawling

Discovering new URLs Identifying duplicates Crawling architecture

2 Web Information Retrieval

3 Web Graph Mining

4 Conclusion

(3)

Web crawling Discovering new URLs

Web Crawlers

crawlers,(Web) spiders,(Web) robots: autonomous user agents that retrieve pages from the Web

Basics of crawling:

1 Start from a given URL or set of URLs

2 Retrieve and process the corresponding page

3 Discover new URLs (cf. next slide)

4 Repeat on each found URL

No real termination condition (virtual unlimited number of Web pages!) Graph-browsingproblem

deep-first: not very adapted, possibility of being lost inrobot traps breadth-first

combination of both: breadth-first with limited-depth deep-first on each discovered website

(4)

Sources of new URLs

From HTML pages:

I hyperlinks<a href="...">...</a>

I media<img src="..."> <embed src="..."> <object data="...">

I frames<frame src="..."> <iframe src="...">

I JavaScript linkswindow.open("...")

I etc.

Other hyperlinked content (e.g., PDF files)

Non-hyperlinked URLs that appear anywhere on the Web (in HTML text, text files, etc.): use regular expressions to extract them

Referrer URLs Sitemaps [sit08]

(5)

Web crawling Discovering new URLs

Scope of a crawler

Web-scale

I The Web is infinite! Avoid robot traps by putting depth or page numberlimits on each Web server

I Focus onimportantpages [APC03] (cf. lecture on the Web graph) Web servers under a list ofDNS domains: easy filtering of URLs

A given topic:focused crawlingtechniques [CvdBD99, DCL⁺00] based on classifiers of Web page content and predictors of the interest of a link.

The national Web (cf.public deposit, national libraries): what is this?

[ACMS02]

A given Web site: what is a Web site? [Sen05]

(6)

A word about hashing

Definition

Ahash functionis a deterministic mathematical function transforming objects (numbers, character strings, binary. . . ) into fixed-size, seemingly random, numbers. The more random the transformation is, the better.

Example

Java hash function for the

String

class:

n−1 i

∑

=0

si

×

31ⁿ⁻ⁱ⁻¹mod 2³² wheres_i is the (Unicode) code of characteri of a strings.

(7)

Web crawling Identifying duplicates

Identification of duplicate Web pages

Problem

Identifying duplicates or near-duplicates on the Web to prevent multiple indexing trivial duplicates: same resource at the samecanonizedURL:

http://example.com:80/toto http://example.com/titi/../toto

exact duplicates: identification byhashing

near-duplicates: (timestamps, tip of the day, etc.) more complex!

(8)

Near-duplicate detection

Edit distance. Count theminimum number of basic modifications(additions or deletions of characters or words, etc.) to obtain a document from another one. Good measure of similarity, and can be computed in O

(

mn

)

wheremandnare the size of the documents. But:does not scaleto a large collection of documents (unreasonable to compute the edit distance for every pair!).

Shingles. Idea: two documents similar if they mostly share the same succession ofk-grams(succession of tokens of lengthk).

Example

I like to watch the sun set with my friend.

My friend and I like to watch the sun set.

S={i like, like to, my friend, set with, sun set, the sun, to watch, watch the, with my}

T ={and i, friend and, i like, like to, my friend, sun set, the sun, to watch, watch the}

(9)

Web crawling Identifying duplicates

Hashing shingles to detect duplicates [BGMZ97]

Similarity:Jaccard coefficienton the set of shingles:

J

(

S,T

) = |

S

∩

T

|

S

∪

T

|

Stillcostly to compute! But can be approximated as follows:

1 ChooseNdifferent hash functions

2 For each hash functionh_i and each set of shinglesS_k={s_k1. . .^skn}, store φik=min_jh_i(s_kj)

3 ApproximateJ(S_k,S_l)as theproportionofφikandφilthat are equal Possibly to repeat in a hierarchical way withsuper-shingles(we are only interested inverysimilar documents)

(10)

Crawling ethics

Standard for robot exclusion: robots.txtat the root of a Web server [Kos94].

User-agent: *

Allow: /searchhistory/

Disallow: /search

Per-page exclusion (de factostandard).

<meta name="ROBOTS" content="NOINDEX,NOFOLLOW">

Per-link exclusion (de factostandard).

<a href="toto.html" rel="nofollow">Toto</a>

AvoidDenial Of Service(DOS), wait 100ms/1s between two repeated requests to the same Web server

(11)

Web crawling Crawling architecture

Parallel processing

Network delays, waits between requests:

Per-server queueof URLs

Parallel processing of requests to different hosts:

I multi-threadedprogramming

I asynchronousinputs and outputs (select, classes from java.util.concurrent): less overhead

Use ofkeep-aliveto reduce connexion overheads

(12)

(13)

Web crawling Crawling architecture

Refreshing URLs

Content on the Webchanges Differentchange rates:

online newspaper main page: every hour or so published article: virtually no change

Continuouscrawling, and identification of change rates foradaptive crawling:

I If-Last-ModifiedHTTP feature (not reliable)

I Identification ofduplicatesin successive request

(14)

Outline

1 Web crawling

2 Web Information Retrieval Text Preprocessing Inverted Index

Answering Keyword Queries

3 Web Graph Mining

4 Conclusion

(15)

Web Information Retrieval

Information Retrieval, Search

Problem

How toindexWeb content so as to answer (keyword-based) queriesefficiently?

Context: set oftext documents

d1 The jaguar is a New World mammal of the Felidae family.

d₂ Jaguar has designed four new engines.

d3 For Jaguar, Atari was keen to use a 68K family device.

d4 The Jacksonville Jaguars are a professional US football team.

d5 Mac OS X Jaguar is available at a price of US $199 for Apple’s new “family pack”.

d6 One such ruling family to incorporate the jaguar into their name is Jaguar Paw.

d₇ It is a big cat.

(16)

Text Preprocessing

Initial textpreprocessingsteps Number of optional steps

Highly depends on theapplication

Highly depends on thedocument language(illustrated with English)

(17)

Web Information Retrieval Text Preprocessing

Language Identification

How to find the language used in a document?

Meta-information about the document: oftennot reliable!

Unambiguousscripts or letters: not very common!

한글 カタカナ

Għarbi þorn

Respectively: Korean Hangul, Japanese Katakana, Maldivian Dhivehi, Maltese, Icelandic

Extension of this:frequent characters, or, better,frequentk-grams Use standard machine learning techniques (classifiers)

(18)

Language Identification

How to find the language used in a document?

Meta-information about the document: oftennot reliable!

Unambiguousscripts or letters: not very common!

한글 カタカナ

Għarbi þorn

Respectively: Korean Hangul, Japanese Katakana, Maldivian Dhivehi, Maltese, Icelandic

Extension of this:frequent characters, or, better,frequentk-grams Use standard machine learning techniques (classifiers)

(19)

Tokenization

Principle

Separate text intotokens(words) Not so easy!

In some languages (Chinese, Japanese), wordsnot separated by whitespace

Dealconsistentlywith acronyms, elisions, numbers, units, URLs, emails, etc.

Compound words: hostname,host-nameandhost name. Break into two tokens or regroup them as one token? In any case, lexicon and linguistic analysis needed! Even more so in other languages as German.

Usually, remove punctuation and normalize case at this point

(20)

Tokenization: Example

d1 the1jaguar2is3a4new5world6mammal7of8the9felidae10family11

d₂ jaguar₁has₂designed₃four₄new₅engines₆

d3 for1jaguar2atari3was4keen5to6use7a868k9family10device11

d4 the1jacksonville2jaguars3are4a5professional6us7football8team9

d₅ mac₁os₂x₃jaguar₄is₅available₆at₇a₈price₉of₁₀us₁₁$199₁₂ for13apple’s14new15family16pack17

d6 one1such2ruling3family4to5incorporate6the7jaguar8into9

their₁₀name₁₁is₁₂jaguar₁₃paw₁₄ d7 it1is2a3big4cat5

(21)

Stemming

Principle

Mergedifferent forms of the same word, or of closely related words, into a single stem

Not in all applications!

Useful for retrieving documents containinggeesewhen searching forgoose Various degreesof stemming

Possibility of building different indexes, with different stemming

(22)

Stemming schemes (1/2)

Morphological stemming.

Removebound morphemesfrom words:

I plural markers

I gender markers

I tense or mood inflections

I etc.

Can be linguisticallyvery complex, cf:

Les poules du couvent couvent.

[The hens of the monastery brood.]

In English, somewhateasy:

I Remove final -s, -’s, -ed, -ing, -er, -est

I Take care of semiregular forms (e.g., -y/-ies)

I Take care of irregular forms (mouse/mice) But still someambiguities: cf stocking, rose

(23)

Stemming schemes (2/2)

Lexical stemming.

Mergelexically relatedterms of various parts of speech, such aspolicy,politics,politicalorpolitician

For English,Porter’s stemming[Por80]; stemuniversityand universaltounivers: not perfect!

Possibility of coupling this withlexiconsto merge (near-)synonyms

Phonetic stemming.

Mergephonetically relatedwords: search despite spelling errors!

For English,Soundex[US 07] stemsRobertandRupertto R163. Verycoarse!

(24)

Stemming Example

d1 the1jaguar2be3a4new5world6mammal7of8the9felidae10family11

d₂ jaguar₁have₂design₃four₄new₅engine₆

d3 for1jaguar2atari3be4keen5to6use7a868k9family10device11

d4 the1jacksonville2jaguar3be4a5professional6us7football8team9

d₅ mac₁os₂x₃jaguar₄be₅available₆at₇a₈price₉of₁₀us₁₁$199₁₂ for13apple14new15family16pack17

d6 one1such2rule3family4to5incorporate6the7jaguar8into9

their₁₀name₁₁be₁₂jaguar₁₃paw₁₄ d7 it1be2a3big4cat5

(25)

Stop Word Removal

Principle

Removeuninformativewords from documents, in particular to lower the cost of storing the index

determiners: a,the,this, etc.

function verbs: be,have,make, etc.

conjunctions: that,and, etc.

etc.

(26)

Stop Word Removal Example

d1 jaguar2new5world6mammal7felidae10family11

d₂ jaguar₁design₃four₄new₅engine₆ d3 jaguar2atari3keen568k9family10device11

d4 jacksonville2jaguar3professional6us7football8team9

d₅ mac₁os₂x₃jaguar₄available₆price₉us₁₁$199₁₂apple₁₄ new15family16pack17

d6 one1such2rule3family4incorporate6jaguar8their10name11

jaguar₁₃paw₁₄ d7 big4cat5

(27)

Web Information Retrieval Inverted Index

Inverted Index construction

After all preprocessing, construction of aninverted index:

Index ofall terms, with the list of documents where this termoccurs Small scale: disk storage, withmemory mapping(cf.

mmap

) techniques;

secondary index for offset of each term in main index

Large scale: distributed on acluster of machines; hashing gives the machine responsible for a given term

Updating the index is costly, so onlybatch operations(not one-by-one addition of term occurrences)

(28)

Inverted Index Example

family d1,d3,d5,d6

football d4

jaguar d1,d2,d3,d4,d5,d6

new d1,d2,d5

rule d6

us d4,d5

world d1

. . . Note:

the length of an inverted (posting) list is highly variable – scanning short lists first is an important optimization.

entriesare homogeneous: this gives much room for compression.

(29)

Storing positions in the index

phrase queries, NEAR operator: need to keepposition informationin the index

just add it in the document list!

family d₁/11,d₃/10,d₅/16,d₆/4 football d4/8

jaguar d1/2,d2/1,d3/2,d4/3,d5/4,d6/8

+

13 new d₁/5,d₂/5,d₅/15

rule d6/3 us d4/7,d5/11 world d₁/6 . . .

⇒

so far, ok forBooleanqueries: find the documents that contain a set of keywords; reject the other.

(30)

TF-IDF Weighting

The inverted is extended by addingTerm Frequency—Inverse Document Frequencyweighting

tfidf

(

t,d

) =

ⁿ^t^,^d

∑t⁰n_t⁰_,_d

·

log

|

D

|

|{

d⁰

∈

D

|

n_t_,_d⁰

>

0

}|

nt,d number of occurrences oftind D set of all documents

Documents (along with weight) are stored indecreasing weight orderin the index

(31)

TF-IDF Weighting Example

family d1/11/.13,d3/10/.13,d6/4/.08,d5/16/.07 football d4/8/.47

jaguar d₁/2/.04,d₂/1/.04,d₃/2/.04,d₄/3/.04,d₆/8

+

13/.04, d5/4/.02

new d2/5/.24,d1/5/.20,d5/15/.10 rule d₆/3/.28

us d4/7/.30,d5/11/.15 world d1/6/.47

. . .

Exercise: take an entry, and check that the tf/idf value is indeed correct (take documents after stop-word removal).

(32)

Answering Boolean Queries

Single keyword query: just consult the index and return the documents in index order.

Boolean multi-keyword query

(jaguarANDnewAND NOTfamily) ORcat

Same way! Retrieve document lists from all keywords and apply adequate set operations:

AND intersection OR union AND NOT difference

Global score: some function of the individual weight (e.g., addition for conjunctive queries)

Position queries: consult the index, and filter by appropriate condition

(33)

Web Information Retrieval Answering Keyword Queries

Answering Top-k Queries

t1AND . . . ANDtn

Problem

Find the top-k results (for some givenk) to the query, without retrieving all documents matching it.

Notations:

s

(

t,d

)

weight oft ind(e.g., tfidf)

g

(

s1,. . .,sn

)

monotonous function that computes the global score (e.g.,

addition)

(34)

The Threshold Algorithm

1 LetRbe the empty list, andm

= +

_∞.

2 For each 1

≤

i

≤

n:

1 Retrieve the documentd⁽ⁱ⁾containing termt_i that has thenext largest s(t_i,d⁽ⁱ⁾).

2 Compute its global scoreg_d(i)=g(s(t₁,d⁽ⁱ⁾),. . .,s(t_n,d⁽ⁱ⁾))by retrieving alls(t_j,d⁽ⁱ⁾)withj6=i.

3 IfRcontains less thankdocuments, or ifg_d(i)is greater than theminimum of the score of documentsinR, addd⁽ⁱ⁾toR.

3 Letm

=

g

(

s

(

t1,d⁽¹⁾

)

,s

(

t2,d⁽²⁾

)

,. . .,s

(

tn,d⁽ⁿ⁾

))

.

4 IfR contains more thank documents, and the minimum of the score of the documents inRisgreater than or equaltom, returnR.

5 Redo step 2.

(35)

Web Information Retrieval Answering Keyword Queries

The TA, by example

q= “newORfamily”, andk

=

3. We use inverted lists sorted on the weight.

family d₁/11/.13,d₃/10/.13,d₆/4/.08,d₅/16/.07 new d2/5/.24,d1/5/.20,d5/15/.10

. . .

Initially,R

= ∅

^and^τ

= +

_∞.

1 d⁽¹⁾is the first entry inL_family, one findss

(

new,d₁

) =

.20; the global score ford1is.13

+

.20

=

.33.

2 Next,i

=

2, and one finds that the global score ford₂is .24.

3 The algorithm quits the loop oniwithR

= h[

d₁, .33

]

,

[

d₂, .24

]i

and τ

=

.13

+

.24

=

.37.

4 We proceed with the loop again, takingd₃with score.13 andd₅with score .17.

[

d5, .17

]

is added toR(at the end) andτis now.10

+

.13

=

.23.

A last loop concludes that the next candidate isd6, with a global score of .08. Then we are done.

(36)

Outline

1 Web crawling

3 Web Graph Mining PageRank HITS Spamdexing

4 Conclusion

(37)

Web Graph Mining

The Web Graph

The World Wide Web seenas a (directed) graph:

Vertices: Web pages Edges: hyperlinks

Same for otherinterlinkedenvironments:

dictionaries encyclopedias scientific publications social networks

(38)

The transition matrix

(g_ij

=

0 if there is no link between pagei andj;

gij

=

_n¹

i otherwise, withni the number of outgoing links of pagei.

G

=







0 1 0 0 0 0 0 0 0 0

0 0 ¹₄ 0 0 ¹₄ ¹₄ 0 ¹₄ 0 0 0 0 ¹₂ ¹₂ 0 0 0 0 0

0 1 0 0 0 0 0 0 0 0

0 0 0 0 0 ¹₂ 0 0 0 ¹₂

1 3

1

3 0 ¹₃ 0 0 0 0 0 0

0 0 0 0 0 ¹₃ 0 ¹₃ 0 ¹₃ 0 ¹₃ 0 0 0 0 0 0 ¹₃ ¹₃ 0 ¹₂ ¹₂ 0 0 0 0 0 0 0

0 0 0 0 1 0 0 0 0 0







(39)

Web Graph Mining PageRank

PageRank (Google’s Ranking [BP98])

Idea

Importantpages are pages pointed to byimportantpages.

PageRank simulates a random walk by iterately computing the PR of each page, represented as a vectorv.

Initially,v is set using a uniform distribution (v

[

i

] =

_|¹

v|).

Definition (Tentative)

Probabilitythat the surfer following therandom walkinGhas arrived on pagei at some distant given point in the future.

pr

(

i

) =

k→+lim_∞

(

G^T

)

^kv

i

wherev is some initial column vector.

(40)

PageRank Iterative Computation

0.100 0.100

0.100

0.100 0.100

0.100

(41)

PageRank Iterative Computation

0.033 0.317

0.075

0.108

0.025 0.058

0.083

0.150

0.117

0.033

(42)

PageRank Iterative Computation

0.036 0.193

0.108

0.163

0.079 0.090

0.074

0.154

0.094

0.008

(43)

PageRank Iterative Computation

0.054 0.212

0.093

0.152

0.048 0.051

0.108

0.149

0.106

0.026

(44)

PageRank Iterative Computation

0.051 0.247

0.078

0.143

0.053 0.062

0.097

0.153

0.099

0.016

(45)

PageRank Iterative Computation

0.048 0.232

0.093

0.156

0.062 0.067

0.087

0.138

0.099

0.018

(46)

PageRank Iterative Computation

0.052 0.226

0.092

0.148

0.058 0.064

0.098

0.146

0.096

0.021

(47)

PageRank Iterative Computation

0.049 0.238

0.088

0.149

0.057 0.063

0.095

0.141

0.099

0.019

(48)

PageRank Iterative Computation

0.050 0.232

0.091

0.149

0.060 0.066

0.094

0.143

0.096

0.019

(49)

PageRank Iterative Computation

0.050 0.233

0.091

0.150

0.058 0.064

0.095

0.142

0.098

0.020

(50)

PageRank Iterative Computation

0.050 0.234

0.090

0.148

0.058 0.065

0.095

0.143

0.097

0.019

(51)

PageRank Iterative Computation

0.049 0.233

0.091

0.149

0.058 0.065

0.095

0.142

0.098

0.019

(52)

PageRank Iterative Computation

0.050 0.233

0.091

0.149

0.058 0.065

0.095

0.143

0.097

0.019

(53)

PageRank Iterative Computation

0.050 0.234

0.091

0.149

0.058 0.065

0.095

0.142

0.097

0.019

(54)

PageRank With Damping

May not always converge, or convergence may not be unique.

To fix this, the random surfer can at each steprandomly jumpto any page of the Web with some probabilityd (1

−

d:damping factor).

pr

(

i

) =

k→+lim_∞

((

1

−

d

)

G^T

+

dU

)

^kv

i

whereUis the matrix with all_N¹ values withNthe number of vertices.

(55)

Using PageRank to Score Query Results

PageRank:globalscore, independent of the query Can be used to raise the weight ofimportantpages:

weight

(

t,d

) =

tfidf

(

t,d

) ×

pr

(

d

)

, This can be directly incorporatedin the index.

(56)

HITS (Kleinberg, [Kle99])

Idea

Two kinds of important pages:hubsandauthorities. Hubs are pages that point to good authorities, whereas authorities are pages that are pointed to by good hubs.

G⁰ transition matrix (with 0 and 1 values) of a subgraph of the Web. We use the following iterative process (starting withaandhvectors of norm 1):

(a:

=

_k ¹

G⁰^ThkG⁰^Th h:

=

_k_G¹0akG⁰a

Convergesunder some technical assumptions toauthorityandhubscores.

(57)

Web Graph Mining HITS

Using HITS to Order Web Query Results

1 Retrieve the setDof Web pagesmatchinga keyword query.

2 Retrieve the setD^∗of Web pages obtained fromDby addingall linked pages, as well as allpages linking topages ofD.

3 Build fromD^∗the correspondingsubgraphG⁰ of the Web graph.

4 Computeiterativelyhubs and authority scores.

5 Sort documents fromDbyauthority scores.

Less efficient than PageRank, becauselocalscores.

(58)

Spamdexing

Definition

Fraudulent techniques that are used by unscrupulous webmasters to artificially raise the visibility of their website to users of search engines

Purpose: attracting visitors to websites to make profit.

Unceasing war betweenspamdexersandsearch engines

(59)

Web Graph Mining Spamdexing

Spamdexing: Lying about the Content

Technique

Putunrelatedterms in:

meta-information (

<meta name="description">

,

<meta name="keywords">

)

text content hidden to the user with JavaScript, CSS, or HTML presentational elements

Countertechnique

Ignoremeta-information Try anddetectinvisible text

(60)

Link Farm Attacks

Technique

Huge number of hosts on the Internet used for the sole purpose ofreferencing each other, without any content in themselves, toraise the importanceof a given website or set of websites.

Countertechnique

Detection of websites withemptyorduplicatecontent

Use of heuristics to discoversubgraphsthat look like link farms

(61)

Web Graph Mining Spamdexing

Link Pollution

Technique

Pollute user-editablewebsites (blogs, wikis) or exploit security bugs to add artificiallinks to websites, in order to raise its importance.

Countertechnique

rel="nofollow"attribute to

<a>

links not validated by a page’s owner

(62)

Outline

1 Web crawling

3 Web Graph Mining

4 Conclusion

(63)

Conclusion

What you should remember

Theinverted indexmodel for efficient answers of keyword-based queries.

Thethreshold algorithmfor retrieving top-k results.

PageRankand its iterative computation.

(64)

References

Specifications

I HTML 4.01,http://www.w3.org/TR/REC-html40/

I HTTP/1.1,http://tools.ietf.org/html/rfc2616

I Robot Exclusion Protocol,

http://www.robotstxt.org/orig.html A book

Mining the Web: Discovering Knowledge from Hypertext Data, Soumen Chakrabarti, Morgan Kaufmann

(65)

Bibliography I

Serge Abiteboul, Grégory Cobena, Julien Masanès, and Gerald Sedrati.

A first experience in archiving the French Web.

InProc. ECDL, Roma, Italie, September 2002.

Serge Abiteboul, Mihai Preda, and Grégory Cobena.

Adaptive on-line page importance computation.

InProc. Intl. World Wide Web Conference (WWW), 2003.

Andrei Z. Broder, Steven C. Glassman, Mark S. Manasse, and Geoffrey Zweig.

Syntactic clustering of the Web.

Computer Networks, 29(8-13):1157–1166, 1997.

Sergey Brin and Lawrence Page.

The anatomy of a large-scale hypertextual Web search engine.

Computer Networks, 30(1–7):107–117, April 1998.

(66)

Soumen Chakrabarti.

Mining the Web: Discovering Knowledge from Hypertext Data.

Morgan Kaufmann, 2003.

Soumen Chakrabarti, Martin van den Berg, and Byron Dom.

Focused crawling: A new approach to topic-specific Web resource discovery.

Computer Networks, 31(11–16):1623–1640, 1999.

Michelangelo Diligenti, Frans Coetzee, Steve Lawrence, C. Lee Giles, and Marco Gori.

Focused crawling using context graphs.

InProc. VLDB, Cairo, Egypt, September 2000.

Jon M. Kleinberg.

Authoritative Sources in a Hyperlinked Environment.

Journal of the ACM, 46(5):604–632, 1999.

(67)

Bibliography III

Martijn Koster.

A standard for robot exclusion.

http://www.robotstxt.org/orig.html, June 1994.

Martin F. Porter.

An algorithm for suffix stripping.

Program, 14(3):130–137, July 1980.

Pierre Senellart.

Identifying Websites with flow simulation.

InProc. ICWE, pages 124–129, Sydney, Australia, July 2005.

sitemaps.org.

Sitemaps XML format.

http://www.sitemaps.org/protocol.php, February 2008.

(68)

US National Archives and Records Administration.

The Soundex indexing system.

http:

//www.archives.gov/genealogy/census/soundex.html, May 2007.