Abstract image of a woman sitting on books and studying on a laptop

Cse535 chapter19-web search

Download as PPT, PDF
Yihong Chen, profile picture
Yihong Chen

This document provides an overview of web search basics and information retrieval. It begins with a brief history of web search engines, from early keyword-based engines to Google pioneering link-based ranking in 1998. It discusses sponsored search and ad placement. The document outlines different types of user needs for search and how far users typically look for results. It also discusses factors in users' evaluation of search engines and results. Other topics covered include the challenges of the web document collection, spam, and techniques for detecting spam like cloaking and content repurposing. The document concludes with a discussion of measuring the size of the web.

TechnologyDesign
1 of 67
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
CSE 435/535 Information Retrieval Fall 2013 Chapter 19: Web Search Basics
Brief (non-technical) history  Early keyword-based engines   Altavista, Excite, Infoseek, Inktomi, ca. 1995-1997 Sponsored search ranking: Goto.com (morphed into Overture.com → Yahoo!)   Your search ranking depended on how much you paid Auction for keywords: casino was expensive!
Brief (non-technical) history  1998+: Link-based ranking pioneered by Google     Blew away all early engines save Inktomi Great user experience in search of a business model Meanwhile Goto/Overture’s annual revenues were nearing $1 billion Result: Google added paid-placement “ads” to the side, independent of search results  Yahoo followed suit, acquiring Overture (for paid placement) and Inktomi (for search)
Ads Algorithmic results.
Web search basics Sponsored Links CG Appliance Express Discount Appliances (650) 756-3931 Same Day Certified Installation www.cgappliance.com San Francisco-Oakland-San Jose, CA User Miele Vacuum Cleaners Miele Vacuums- Complete Selection Free Shipping! www.vacuums.com Miele Vacuum Cleaners Miele-Free Air shipping! All models. Helpful advice. www.best-vacuum.com Web Results 1 - 10 of about 7,310,000 for miele. (0.12 seconds) Miele, Inc -- Anything else is a compromise Web spider At the heart of your home, Appliances by Miele. ... USA. to miele.com. Residential Appliances. Vacuum Cleaners. Dishwashers. Cooking Appliances. Steam Oven. Coffee System ... www.miele.com/ - 20k - Cached - Similar pages Miele Welcome to Miele, the home of the very best appliances and kitchens in the world. www.miele.co.uk/ - 3k - Cached - Similar pages Miele - Deutscher Hersteller von Einbaugeräten, Hausgeräten ... - [ Translate this page ] Das Portal zum Thema Essen & Geniessen online unter www.zu-tisch.de. Miele weltweit ...ein Leben lang. ... Wählen Sie die Miele Vertretung Ihres Landes. www.miele.de/ - 10k - Cached - Similar pages Herzlich willkommen bei Miele Österreich - [ Translate this page ] Herzlich willkommen bei Miele Österreich Wenn Sie nicht automatisch weitergeleitet werden, klicken Sie bitte hier! HAUSHALTSGERÄTE ... www.miele.at/ - 3k - Cached - Similar pages Search Indexer The Web Indexes Ad indexes
User Needs  Need [Brod02, RL04]    Informational – want to learn about something (~40% / 65%) Low hemoglobin Navigational – want to go to that page (~25% / 15%) United Airlines Transactional – want to do something (web-mediated) (~35% / 20%)  Access a service Seattle weather Mars surface images  Downloads   Shop Canon S410 Gray areas   Car rental Brasil Find a good hub Exploratory search “see what’s there”
How far do people look for results? (Source: iprospect.com WhitePaper_2006_SearchEngineUserBehavior.pdf)
Users’ empirical evaluation of results  Quality of pages varies widely   Relevance is not enough Other desirable qualities (non IR!!)     Precision vs. recall   On the web, recall seldom matters What matters   Precision at 1? Precision above the fold? Comprehensiveness – must be able to deal with obscure queries   Content: Trustworthy, diverse, non-duplicated, well maintained Web readability: display correctly & fast No annoyances: pop-ups, etc Recall matters when the number of matches is very small User perceptions may be unscientific, but are significant over a large aggregate
Users’ empirical evaluation of engines      Relevance and validity of results UI – Simple, no clutter, error tolerant Trust – Results are objective Coverage of topics for polysemic queries Pre/Post process tools provided     Mitigate user errors (auto spell check, search assist,…) Explicit: Search within results, more like this, refine ... Anticipative: related searches Deal with idiosyncrasies  Web specific vocabulary    Impact on stemming, spell-check, etc Web addresses typed in the search box …
The Web document collection       The Web  No design/co-ordination Distributed content creation, linking, democratization of publishing Content includes truth, lies, obsolete information, contradictions … Unstructured (text, html, …), semistructured (XML, annotated photos), structured (Databases)… Scale much larger than previous text collections … but corporate records are catching up Growth – slowed down from initial “volume doubling every few months” but still expanding Content can be dynamically generated
Spam Search Engine Optimization
The trouble with sponsored search …   It costs money. What’s the alternative? Search Engine Optimization:      “Tuning” your web page to rank highly in the algorithmic search results for select keywords Alternative to paying for placement Thus, intrinsically a marketing function Performed by companies, webmasters and consultants (“Search engine optimizers”) for their clients Some perfectly legitimate, some very shady
Simplest forms  First generation engines relied heavily on tf/idf   The top-ranked pages for the query maui resort were the ones containing the most maui’s and resort’s SEOs responded with dense repetitions of chosen terms   e.g., maui resort maui resort maui resort Often, the repetitions would be in the same color as the background of the web page   Repeated terms got indexed by crawlers But not visible to humans on browsers Pure word density cannot be trusted as an IR signal
Variants of keyword stuffing   Misleading meta-tags, excessive repetition Hidden text with colors, style sheet tricks, etc. Meta-Tags = “… London hotels, hotel, holiday inn, hilton, discount, booking, reservation, sex, mp3, britney spears, viagra, …”
Search engine optimization (Spam)  Motives    Operators     Commercial, political, religious, lobbies Promotion funded by advertising budget Contractors (Search Engine Optimizers) for lobbies, companies Web masters Hosting services Forums  E.g., Web master world ( www.webmasterworld.com )   Search engine specific tricks Discussions about academic papers 
Synthetic content Monetization Random words Well-formed sentences stitched together Links to keep crawlers going
Features identifying synthetic content  Average word length   Word frequency distribution   Certain words (“the”, “a”, …) appear more often than others N-gram frequency distribution   The mean word length for English prose is about 5 characters; but longer for some forms of keyword stuffing Some words are more likely to occur next to each other than others Grammatical well-formedness  Natural-language parsing is expensive
Really good synthetic content “Nigritude Ultramarine”: An SEO competition Links to keep crawlers going Grammatically well-formed but meaningless sentences
Content “repurposing”  Content repurposing: The practice of incorporating all or portions of other (unaffiliated) web pages    A “convenient” way to machine generate pages that contain human-authored content Not even necessarily illegal … Two flavors:   Incorporate large portions of a single page Incorporate snippets of multiple pages
Example of page-level content “repurposing”
Example of phrase-level content “repurposing” How is monetization taking place?
Techniques for detecting content repurposing  Single-page flavor: Cluster pages into equivalence classes of very similar pages    If most pages on a site a very similar to pages on other sites, raise a red flag (There are legitimate replicated sites; e.g. mirrors of Linux man pages) Many-snippets flavor: Test if page consists mostly of phrases that also occur somewhere else   Computationally hard problem Probabilistic technique that makes it tractable (Fetterly et al SIGIR 2005 paper)
Cloaking   Serve fake content to search engine spider DNS cloaking: Switch IP address. Impersonate Y SPAM Is this a Search Engine spider? Cloaking N Real Doc
More spam techniques  Doorway pages   Link spamming     Pages optimized for a single keyword that re-direct to the real target page links between pages that are present for reasons other than merit Mutual admiration societies, hidden links, awards – more on these later Domain flooding: numerous domains that point or re-direct to a target page Robots  Fake query stream – rank checking programs   “Curve-fit” ranking programs of search engines Millions of submissions via Add-Url
The war against spam  Quality signals - Prefer authoritative pages based on:  Votes from authors (linkage signals) Votes from users (usage signals)        Anti robot test , i.e. not submitted by a robot Limits on meta-keywords Robust link analysis   Ignore statistically implausible linkage (or text) Use link analysis to detect spammers (guilt by association)   Training set based on known spam Family friendly filters  Policing of URL submissions  Spam recognition by machine learning Linguistic analysis, general classification techniques, etc. For images: flesh tone detectors, source text analysis, etc. Editorial intervention     Blacklists Top queries audited Complaints addressed Suspect pattern detection
More on spam  Web search engines have policies on SEO practices they tolerate/block     http://help.yahoo.com/help/us/ysearch/index.html http://www.google.com/intl/en/webmasters/ Adversarial IR: the unending (technical) battle between SEO’s and web search engines Research http://airweb.cse.lehigh.edu/
Size of the web
What is the size of the web ?  Issues  The web is really infinite      Dynamic content, e.g., calendar Soft 404: www.yahoo.com/<anything> is a valid page Static web contains syntactic duplication, mostly due to mirroring (~30%) Some servers are seldom connected Who cares?    Media, and consequently the user Engine design Engine crawl policy. Impact on recall.
What can we attempt to measure? The relative sizes of search engines    The notion of a page being indexed is still reasonably well defined. Already there are problems   Document extension: e.g. engines index pages not yet crawled, by indexing anchortext. Document restriction: All engines restrict what is indexed (first n words, only relevant words, etc.) The coverage of a search engine relative to another particular crawling process. 
Graph Structure in the Web http://www9.org/w9cdrom/160/160.html
Zipf’s Law on the Web  Number of in-links/out-links to/from a page has a Zipfian distribution.   Length of web pages has a Zipfian distribution.   A large number of docs have the same number of in-links/out-links to/from a page Certain page lengths are most common; i.e. a huge number of docs with the same page length. Number of hits to a web page has a Zipfian distribution.  A large number of pages will generate the same number of hits; dramatic drop-off in number of pages as the above frequency reduces
New definition?  (IQ is whatever the IQ tests measure.)   Different engines have different preferences   The statically indexable web is whatever search engines index. max url depth, max count/host, anti-spam rules, priority rules, etc. Different engines index different things under the same URL:  frames, meta-keywords, document restrictions, document extensions, ...
Relative Size from Overlap Given two engines A and B Sample URLs randomly from A Check if contained in B and vice versa A∩ B A∩ B = A∩ B = (1/2) * Size A (1/6) * Size B (1/2)*Size A = (1/6)*Size B ∴ Size A / Size B = (1/6)/(1/2) = 1/3 Each test involves: (i) Sampling (ii) Checking
Sampling URLs  Ideal strategy: Generate a random URL and check for containment in each index.  Problem: Random URLs are hard to find! Enough to generate a random URL contained in a given Engine. Approach 1: Generate a random URL contained in a given engine     Pick a random query from search log; pick random page returned as search result Suffices for the estimation of relative size Approach 2: Random walks / IP addresses  In theory: might give us a true estimate of the size of the web (as opposed to just relative sizes of indexes)
Statistical methods  Approach 1    *** Random queries: preferred method Random searches Approach 2   Random IP addresses Random walks
Random URLs from random queries  Generate random query: how?   Not an English Lexicon: 400,000+ words from a web crawl dictionary Conjunctive Queries: w1 and w2 e.g., vocalists AND rsi     Get 100 result URLs from engine A Choose a random URL as the candidate to check for presence in engine B (how do you do this?) This distribution induces a probability weight W(p) for each page. Conjecture: W(SEA) / W(SEB) ~ |SEA| / |SEB|
Query Based Checking  Strong Query to check whether an engine B has a document D (corresponding to random URL):     Download D. Get list of words. Use 8 low frequency words as AND query to B Check if D is present in result set. Problems:      Near duplicates (not D, but something identical to D) Frames Redirects Engine time-outs Is 8-word query good enough?
Advantages & disadvantages   Statistically sound under the induced weight. Biases induced by random query  Query Bias: Favors content-rich pages in the language(s) of the lexicon    Ranking Bias: Solution: Use conjunctive queries & fetch all Checking Bias: Duplicates, impoverished pages omitted Document or query restriction bias: engine might not deal properly with 8 words conjunctive query   Malicious Bias: Sabotage by engine Operational Problems: Time-outs, failures, engine inconsistencies, index modification.
Alt 1: Random searches   These are real queries based on real users, rather than “random queries” Choose random searches extracted from a local log [Lawrence & Giles 97] or build “random searches” [Notess]    Use only queries with small results sets. Count normalized URLs in result sets. Use ratio statistics
Advantages & disadvantages  Advantage   Might be a better reflection of the human perception of coverage Issues    Samples are correlated with source of log Duplicates Technical statistical problems (must have nonzero results, ratio average not statistically sound)
Random searches    575 & 1050 queries from the NEC RI employee logs 6 Engines in 1998, 11 in 1999 Implementation:  Restricted to queries with < 600 results in total  Counted URLs from each engine after verifying query match  Computed size ratio & overlap for individual queries  Estimated index size ratio & overlap by averaging over all queries
Queries from Lawrence and Giles study           adaptive access control neighborhood preservation topographic hamiltonian structures right linear grammar pulse width modulation neural unbalanced prior probabilities ranked assignment method internet explorer favourites importing karvel thornber zili liu           softmax activation function bose multidimensional system theory gamma mlp dvi2pdf john oliensis rieke spikes exploring neural video watermarking counterpropagation network fat shattering dimension abelson amorphous computing
Alt 2: Random IP addresses   Generate random IP addresses Find a web server at the given address   If there’s one Collect all pages from server  From this, choose a page at random
Random IP addresses  HTTP requests to random IP addresses    Ignored: empty or authorization required or excluded [Lawr99] Estimated 2.8 million IP addresses running crawlable web servers (16 million total) from observing 2500 servers. OCLC using IP sampling found 8.7 M hosts in 2001   Netcraft [Netc02] accessed 37.2 million hosts in July 2002 [Lawr99] exhaustively crawled 2500 servers and extrapolated   Estimated size of the web to be 800 million Estimated use of metadata descriptors:  Meta tags (keywords, description) in 34% of home pages, Dublin core metadata in 0.3%
Advantages & disadvantages  Advantages    Clean statistics Independent of crawling strategies Disadvantages    Doesn’t deal with duplication Many hosts might share one IP, or not accept requests No guarantee all pages are linked to root page.   Power law for # pages/hosts generates bias towards sites with few pages.   Eg: employee pages But bias can be accurately quantified IF underlying distribution understood Potentially influenced by spamming (multiple IP’s for same server to avoid IP block)
Alt 3: Random walks   View the Web as a directed graph Build a random walk on this graph  Includes various “jump” rules back to visited sites    Converges to a stationary distribution      Does not get stuck in spider traps! Can follow all links! Must assume graph is finite and independent of the walk. Conditions are not satisfied (cookie crumbs, flooding) Time to convergence not really known Sample from stationary distribution of walk Use the “strong query” method to check coverage by SE
Advantages & disadvantages  Advantages    “Statistically clean” method at least in theory! Could work even for infinite web (assuming convergence) under certain metrics. Disadvantages    List of seeds is a problem. Practical approximation might not be valid. Non-uniform distribution  Subject to link spamming
Conclusions     No sampling solution is perfect. Lots of new ideas ... ....but the problem is getting harder Quantitative studies are fascinating and a good research problem
Duplicate detection
Duplicate documents   The web is full of duplicated content Strict duplicate detection = exact match   Not as common But many, many cases of near duplicates  E.g., Last modified date the only difference between two copies of a page
Duplicate/Near-Duplicate Detection   Duplication: Exact match can be detected with fingerprints (64-bit), if fingerprints match, check further Near-Duplication: Approximate match  Overview   Compute syntactic similarity with an edit-distance measure Use similarity threshold to detect near-duplicates   E.g., Similarity > 80% => Documents are “near duplicates” Not transitive though sometimes used transitively
Computing Similarity   Features:  Segments of a document (natural or artificial breakpoints)  Shingles (Word N-Grams)  a rose is a rose is a rose → a_rose_is_a rose_is_a_rose is_a_rose_is a_rose_is_a Similarity Measure between two docs (= sets of shingles)  Set intersection  Specifically (Size_of_Intersection / Size_of_Union)
Shingles + Set Intersection Computing exact set intersection of shingles between all pairs of documents is expensive/intractable   Approximate using a cleverly chosen subset of shingles from each (a sketch) Estimate (size_of_intersection / size_of_union) based on a short sketch  Doc Doc A A Shingle set A Doc Doc B B Shingle set A Sketch A Jaccard Sketch A
Sketch of a document  Create a “sketch vector” (of size ~200) for each document   Documents that share ≥ t (say 80%) corresponding vector elements are near duplicates For doc D, sketchD[ i ] is as follows:    Let f map all shingles in the universe to 0..2m (e.g., f = fingerprinting) Let πi be a random permutation on 0..2m Pick MIN {πi(f(s))} over all shingles s in D
Computing Sketch[i] for Doc1 Document 1 264 Start with 64-bit f(shingles) 264 Permute on the number line 2 64 264 with πi Pick the min value
Test if Doc1.Sketch[i] = Doc2.Sketch[i] Document 2 Document 1 264 264 264 264 A 264 264 2 64 B Are these equal? Test for 200 random permutations: π1, π2,… π200 264
However… Document 2 Document 1 264 264 264 264 A 264 264 B 264 264 A = B iff the shingle with the MIN value in the union of Doc1 and Doc2 is common to both (i.e., lies in the intersection) Why? Claim: This happens with probability Size_of_intersection / Size_of_union
Set Similarity of sets Ci , Cj Jaccard(Ci , C j ) =   Ci  C j Ci  C j View sets as columns of a matrix A; one row for each element in the universe. aij = 1 indicates presence of item i in set j C1 C2 Example 0 1 1 0 1 1 0 1 0 1 Jaccard(C 1 ,C 2 ) = 2/5 = 0.4
Key Observation  For columns Ci, Cj, four types of rows Ci A B C D   1 1 0 0 Cj 1 0 1 0 Overload notation: A = # of rows of type A Claim A Jaccard(Ci , C j ) = A+B+C
“Min” Hashing   Randomly permute rows Hash h(Ci) = index of first row with 1 in column Ci  Surprising Property  Why? P [ h(Ci ) = h(C j ) ] = Jaccard ( Ci , C j )    Both are A/(A+B+C) Min: Look down columns Ci, Cj until first non-Type-D row h(Ci) = h(Cj)  type A row  Because we chose a random permutation, prob of this occurring is A/(A+B+C)
Min-Hash sketches   Pick P random row permutations MinHash sketch SketchD = list of P (usually, 200) indexes (i.e. integers) of first rows with 1 in column C  Similarity of signatures   Let sim[sketch(Ci),sketch(Cj)] = fraction of permutations where MinHash values agree Observe E[sim(sig(Ci),sig(Cj))] = Jaccard(Ci,Cj)
Example R1 R2 R3 R4 R5 C1 1 0 1 1 0 C2 0 1 0 0 1 C3 1 1 0 1 0 Signatures (composed of min) S1 S2 S3 Perm 1 = (12345) 1 2 1 Perm 2 = (54321) 4 5 4 Perm 3 = (34512) 3 5 4 Similarities 1-2 1-3 2-3 Col-Col 0.00 0.50 0.25 Sig-Sig 0.00 0.67 0.00 Good estimate
Implementation Trick   Permuting universe even once is prohibitive Row Hashing    Pick P hash functions hk: {1,…,n}{1,…,O(n)} Ordering under hk gives random permutation of rows One-pass Implementation  For each Ci and hk, keep “slot” for min-hash value  Initialize all slot(Ci,hk) to infinity  Scan rows in arbitrary order looking for 1’s  Suppose row Rj has 1 in column Ci  For each hk,
Example R1 R2 R3 R4 R5 C1 1 0 1 1 0 C2 0 1 1 0 1 h(x) = x mod 5 g(x) = 2x+1 mod 5 h(1) = 1 g(1) = 3 C1 slots C2 slots 1 3 - h(2) = 2 g(2) = 0 1 3 2 0 h(3) = 3 g(3) = 2 1 2 2 0 h(4) = 4 g(4) = 4 h(5) = 0 g(5) = 1 1 2 1 2 2 0 0 0
Comparing Signatures   Signature Matrix S  Rows = Hash Functions  Columns = Columns  Entries = Signatures Can compute – Pair-wise similarity of any pair of signature columns
All signature pairs   Now we have an extremely efficient method for estimating a Jaccard coefficient for a single pair of documents. But we still have to estimate N2 coefficients where N is the number of web pages.    Still slow One solution: locality sensitive hashing (LSH) Another solution: sorting (Henzinger 2006)
More resources  IIR Chapter 19

More Related Content

PPTX
Search Engine Optimization Concepts - The Future of Search
Geromme Talampas
 
PPT
INTRODUCTION TO INFORMATION RETRIVAL
sathish sak
 
PPT
How to Build a Prospecting Database from Scratch 2010
prospectstogo
 
PPT
Search Engine Optimization Ranking Factors
Gerry Grant
 
PDF
Internet Intelligence
robcrayford
 
PPTX
Concordia Universtiy Seo Services Seminar
Greg Bennett
 
PPTX
Local SEO Basics
Inflow
 
PPTX
Competitive Intelligence For Search Marketers
Igal Stolpner
 
Search Engine Optimization Concepts - The Future of Search
Geromme Talampas
 
INTRODUCTION TO INFORMATION RETRIVAL
sathish sak
 
How to Build a Prospecting Database from Scratch 2010
prospectstogo
 
Search Engine Optimization Ranking Factors
Gerry Grant
 
Internet Intelligence
robcrayford
 
Concordia Universtiy Seo Services Seminar
Greg Bennett
 
Local SEO Basics
Inflow
 
Competitive Intelligence For Search Marketers
Igal Stolpner
 

What's hot (20)

PPT
Search Engine Marketing - Search Engines Background and SEO Introduction
Jon Rognerud Chaosmap Digital
 
PPT
Seo & Internet Marketing
alafrancis
 
DOC
Seo Manual
imgaurav16
 
PPTX
Introduction to SEO
Rand Fishkin
 
PDF
What you need to know about seo for Rocks Digital Conference by Barry Schwartz
Barry Schwartz
 
PDF
SEO Workshop
Internet República
 
PPT
Getting the most out of your web site.
Stephen Shaw
 
PPTX
Onsite seo from the wizard of moz
Umberto Tessitore
 
PPT
SEO: search Engine Optimization
phoolchand yadav
 
PPTX
Acting local in a global world (local SEO)
Similarweb
 
PPT
Digital Discoveries: Strategies to End 2013 with Online Business Success
tbkCreative
 
DOCX
Collins book (1) (1)
Simcol Digital Advertising Media
 
PPT
Seo ppt
kalaiyarasi11
 
PPT
Emarketing
Saleh Shneiwer
 
PPT
Seo
Eguardian India
 
PPTX
Emarketing [repaired]
stephen Ward
 
PPT
Web analytics
racerunner
 
PPTX
Competitive Intelligence For Competitive SEOs
JoelBondorowsky
 
PPT
Emarketing
bavalekar
 
PPT
John Jantsch's Keynote Presentation at the eMarketing For Entrepreneurs Confe...
Corporate College
 
Search Engine Marketing - Search Engines Background and SEO Introduction
Jon Rognerud Chaosmap Digital
 
Seo & Internet Marketing
alafrancis
 
Seo Manual
imgaurav16
 
Introduction to SEO
Rand Fishkin
 
What you need to know about seo for Rocks Digital Conference by Barry Schwartz
Barry Schwartz
 
SEO Workshop
Internet República
 
Getting the most out of your web site.
Stephen Shaw
 
Onsite seo from the wizard of moz
Umberto Tessitore
 
SEO: search Engine Optimization
phoolchand yadav
 
Acting local in a global world (local SEO)
Similarweb
 
Digital Discoveries: Strategies to End 2013 with Online Business Success
tbkCreative
 
Collins book (1) (1)
Simcol Digital Advertising Media
 
Seo ppt
kalaiyarasi11
 
Emarketing
Saleh Shneiwer
 
Seo
Eguardian India
 
Emarketing [repaired]
stephen Ward
 
Web analytics
racerunner
 
Competitive Intelligence For Competitive SEOs
JoelBondorowsky
 
Emarketing
bavalekar
 
John Jantsch's Keynote Presentation at the eMarketing For Entrepreneurs Confe...
Corporate College
 
Ad

Similar to Cse535 chapter19-web search (20)

PPTX
SEO-HIGH TRAFFIC ROUTING
BUDNET
 
PPT
(2004) Using internet search engines to market your services
OfficeFinder Network / TrueView360s
 
PDF
How FINN became somewhat search engine friendly @ Oslo SEO meetup 2018
Henning Spjelkavik
 
PPT
Introduction to search_marketing
Bill Hunt
 
PPT
Search Enginesv2
athiracyborg
 
PDF
Digital Marketing
Digipro India
 
PPT
What is SEO? and The Seven Steps to SEO in easy way
ddarkdarkdark755
 
PPT
seo (1).ppt
ssuser4ab089
 
PPT
seo.ppt
ssuser53c1282
 
PPT
Seo Beginners Slide Show
Tin180 VietNam
 
PPT
Vertical Search - challenges and opportunities
Andy Black
 
PPTX
SEO
Binary Web Solutions (I) Pvt Ltd
 
PPTX
17 00 distil rami
Property Portal Watch
 
PPT
Free Traffic Seo Smo 101 (Search Engine Social Media Optimization) Present...
jward5519
 
PPT
SEO Secrets for Online Retailers Webinar
getelastic
 
PPT
BEST SEO TIPS - Tricks SEO
vickybish
 
PPT
Vikram seo ppt
vickybish
 
PPT
SEO PPT
Prateek Teotia
 
PDF
Seo
pardeepkaur60
 
PPTX
1629885596 seoppt1 130303013241-phpapp02
DpSaini1
 
SEO-HIGH TRAFFIC ROUTING
BUDNET
 
(2004) Using internet search engines to market your services
OfficeFinder Network / TrueView360s
 
How FINN became somewhat search engine friendly @ Oslo SEO meetup 2018
Henning Spjelkavik
 
Introduction to search_marketing
Bill Hunt
 
Search Enginesv2
athiracyborg
 
Digital Marketing
Digipro India
 
What is SEO? and The Seven Steps to SEO in easy way
ddarkdarkdark755
 
seo (1).ppt
ssuser4ab089
 
seo.ppt
ssuser53c1282
 
Seo Beginners Slide Show
Tin180 VietNam
 
Vertical Search - challenges and opportunities
Andy Black
 
SEO
Binary Web Solutions (I) Pvt Ltd
 
17 00 distil rami
Property Portal Watch
 
Free Traffic Seo Smo 101 (Search Engine Social Media Optimization) Present...
jward5519
 
SEO Secrets for Online Retailers Webinar
getelastic
 
BEST SEO TIPS - Tricks SEO
vickybish
 
Vikram seo ppt
vickybish
 
SEO PPT
Prateek Teotia
 
Seo
pardeepkaur60
 
1629885596 seoppt1 130303013241-phpapp02
DpSaini1
 
Ad

More from Yihong Chen (9)

PDF
0.9
Yihong Chen
 
PDF
0.5
Yihong Chen
 
PDF
0.3
Yihong Chen
 
PDF
0 0.3
Yihong Chen
 
PDF
Contributer
Yihong Chen
 
PDF
上海金炳2011年一季度金价预测报告
Yihong Chen
 
PDF
上海金炳2011年金价预测报告
Yihong Chen
 
PDF
上海金炳2011年2月金价预测报告
Yihong Chen
 
PDF
上海金炳2011年一季度金价预测报告BAK
Yihong Chen
 
0.9
Yihong Chen
 
0.5
Yihong Chen
 
0.3
Yihong Chen
 
0 0.3
Yihong Chen
 
Contributer
Yihong Chen
 
上海金炳2011年一季度金价预测报告
Yihong Chen
 
上海金炳2011年金价预测报告
Yihong Chen
 
上海金炳2011年2月金价预测报告
Yihong Chen
 
上海金炳2011年一季度金价预测报告BAK
Yihong Chen
 

Recently uploaded (20)

PDF
Gestión Unificada de los Riegos Externos
Cristian Garcia G.
 
PDF
EGCB_Solar_Project_Presentation_and Finalcial Analysis.pdf
TaukirIslam
 
PPT
Overviiew on Intellectual property right
Ritu Maity
 
PDF
NewMind AI Journal Monthly Chronicles - August 2025
NewMind AI
 
PDF
eBook Outline_ AI in Cybersecurity – The Future of Digital Defense.pdf
samalmaheswar155
 
PDF
Secure Java Applications against Quantum Threats
Ana-Maria Mihalceanu
 
PDF
Optimizing bioinformatics applications: a novel approach with human protein d...
IAESIJAI
 
PDF
ment.tech-How to Develop an AI Agent Healthcare App like Sully AI (1).pdf
thomasnova26
 
PPTX
Report in SIP_Distance_Learning_Technology_Impact.pptx
KahelCaponesPamitang
 
PDF
The Basics of Artificial Intelligence - Understanding the Key Concepts and Te...
FODUU
 
PDF
Rooftops detection with YOLOv8 from aerial imagery and a brief review on roof...
IAESIJAI
 
PPTX
From Curiosity to ROI — Cost-Benefit Analysis of Agentic Automation [3/6]
suhanisingh58689
 
PDF
Child-friendly e-learning for artificial intelligence education in Indonesia:...
IAESIJAI
 
PPTX
Blending method and technology for hydrogen.pptx
PradipChanda5
 
PDF
Examining Bias in AI Generated News Content.pdf
Government
 
PDF
“Introduction to Designing with AI Agents,” a Presentation from Amazon Web Se...
Edge AI and Vision Alliance
 
PDF
Slides World Game (s) Great Redesign Eco Economic Epochs.pdf
Steven McGee
 
PDF
Revolutionizing recommendations a survey: a comprehensive exploration of mode...
IAESIJAI
 
PPTX
AQUEEL MUSHTAQUE FAKIH COMPUTER CENTER .
nachanmasarrat13
 
PPTX
Presentation - Principles of Instructional Design.pptx
Ntokozo Mhlongo
 
Gestión Unificada de los Riegos Externos
Cristian Garcia G.
 
EGCB_Solar_Project_Presentation_and Finalcial Analysis.pdf
TaukirIslam
 
Overviiew on Intellectual property right
Ritu Maity
 
NewMind AI Journal Monthly Chronicles - August 2025
NewMind AI
 
eBook Outline_ AI in Cybersecurity – The Future of Digital Defense.pdf
samalmaheswar155
 
Secure Java Applications against Quantum Threats
Ana-Maria Mihalceanu
 
Optimizing bioinformatics applications: a novel approach with human protein d...
IAESIJAI
 
ment.tech-How to Develop an AI Agent Healthcare App like Sully AI (1).pdf
thomasnova26
 
Report in SIP_Distance_Learning_Technology_Impact.pptx
KahelCaponesPamitang
 
The Basics of Artificial Intelligence - Understanding the Key Concepts and Te...
FODUU
 
Rooftops detection with YOLOv8 from aerial imagery and a brief review on roof...
IAESIJAI
 
From Curiosity to ROI — Cost-Benefit Analysis of Agentic Automation [3/6]
suhanisingh58689
 
Child-friendly e-learning for artificial intelligence education in Indonesia:...
IAESIJAI
 
Blending method and technology for hydrogen.pptx
PradipChanda5
 
Examining Bias in AI Generated News Content.pdf
Government
 
“Introduction to Designing with AI Agents,” a Presentation from Amazon Web Se...
Edge AI and Vision Alliance
 
Slides World Game (s) Great Redesign Eco Economic Epochs.pdf
Steven McGee
 
Revolutionizing recommendations a survey: a comprehensive exploration of mode...
IAESIJAI
 
AQUEEL MUSHTAQUE FAKIH COMPUTER CENTER .
nachanmasarrat13
 
Presentation - Principles of Instructional Design.pptx
Ntokozo Mhlongo
 

Cse535 chapter19-web search

  • 1. CSE 435/535 Information Retrieval Fall 2013 Chapter 19: Web Search Basics
  • 2. Brief (non-technical) history  Early keyword-based engines   Altavista, Excite, Infoseek, Inktomi, ca. 1995-1997 Sponsored search ranking: Goto.com (morphed into Overture.com → Yahoo!)   Your search ranking depended on how much you paid Auction for keywords: casino was expensive!
  • 3. Brief (non-technical) history  1998+: Link-based ranking pioneered by Google     Blew away all early engines save Inktomi Great user experience in search of a business model Meanwhile Goto/Overture’s annual revenues were nearing $1 billion Result: Google added paid-placement “ads” to the side, independent of search results  Yahoo followed suit, acquiring Overture (for paid placement) and Inktomi (for search)
  • 5. Web search basics Sponsored Links CG Appliance Express Discount Appliances (650) 756-3931 Same Day Certified Installation www.cgappliance.com San Francisco-Oakland-San Jose, CA User Miele Vacuum Cleaners Miele Vacuums- Complete Selection Free Shipping! www.vacuums.com Miele Vacuum Cleaners Miele-Free Air shipping! All models. Helpful advice. www.best-vacuum.com Web Results 1 - 10 of about 7,310,000 for miele. (0.12 seconds) Miele, Inc -- Anything else is a compromise Web spider At the heart of your home, Appliances by Miele. ... USA. to miele.com. Residential Appliances. Vacuum Cleaners. Dishwashers. Cooking Appliances. Steam Oven. Coffee System ... www.miele.com/ - 20k - Cached - Similar pages Miele Welcome to Miele, the home of the very best appliances and kitchens in the world. www.miele.co.uk/ - 3k - Cached - Similar pages Miele - Deutscher Hersteller von Einbaugeräten, Hausgeräten ... - [ Translate this page ] Das Portal zum Thema Essen & Geniessen online unter www.zu-tisch.de. Miele weltweit ...ein Leben lang. ... Wählen Sie die Miele Vertretung Ihres Landes. www.miele.de/ - 10k - Cached - Similar pages Herzlich willkommen bei Miele Österreich - [ Translate this page ] Herzlich willkommen bei Miele Österreich Wenn Sie nicht automatisch weitergeleitet werden, klicken Sie bitte hier! HAUSHALTSGERÄTE ... www.miele.at/ - 3k - Cached - Similar pages Search Indexer The Web Indexes Ad indexes
  • 6. User Needs  Need [Brod02, RL04]    Informational – want to learn about something (~40% / 65%) Low hemoglobin Navigational – want to go to that page (~25% / 15%) United Airlines Transactional – want to do something (web-mediated) (~35% / 20%)  Access a service Seattle weather Mars surface images  Downloads   Shop Canon S410 Gray areas   Car rental Brasil Find a good hub Exploratory search “see what’s there”
  • 7. How far do people look for results? (Source: iprospect.com WhitePaper_2006_SearchEngineUserBehavior.pdf)
  • 8. Users’ empirical evaluation of results  Quality of pages varies widely   Relevance is not enough Other desirable qualities (non IR!!)     Precision vs. recall   On the web, recall seldom matters What matters   Precision at 1? Precision above the fold? Comprehensiveness – must be able to deal with obscure queries   Content: Trustworthy, diverse, non-duplicated, well maintained Web readability: display correctly & fast No annoyances: pop-ups, etc Recall matters when the number of matches is very small User perceptions may be unscientific, but are significant over a large aggregate
  • 9. Users’ empirical evaluation of engines      Relevance and validity of results UI – Simple, no clutter, error tolerant Trust – Results are objective Coverage of topics for polysemic queries Pre/Post process tools provided     Mitigate user errors (auto spell check, search assist,…) Explicit: Search within results, more like this, refine ... Anticipative: related searches Deal with idiosyncrasies  Web specific vocabulary    Impact on stemming, spell-check, etc Web addresses typed in the search box …
  • 10. The Web document collection       The Web  No design/co-ordination Distributed content creation, linking, democratization of publishing Content includes truth, lies, obsolete information, contradictions … Unstructured (text, html, …), semistructured (XML, annotated photos), structured (Databases)… Scale much larger than previous text collections … but corporate records are catching up Growth – slowed down from initial “volume doubling every few months” but still expanding Content can be dynamically generated
  • 12. The trouble with sponsored search …   It costs money. What’s the alternative? Search Engine Optimization:      “Tuning” your web page to rank highly in the algorithmic search results for select keywords Alternative to paying for placement Thus, intrinsically a marketing function Performed by companies, webmasters and consultants (“Search engine optimizers”) for their clients Some perfectly legitimate, some very shady
  • 13. Simplest forms  First generation engines relied heavily on tf/idf   The top-ranked pages for the query maui resort were the ones containing the most maui’s and resort’s SEOs responded with dense repetitions of chosen terms   e.g., maui resort maui resort maui resort Often, the repetitions would be in the same color as the background of the web page   Repeated terms got indexed by crawlers But not visible to humans on browsers Pure word density cannot be trusted as an IR signal
  • 14. Variants of keyword stuffing   Misleading meta-tags, excessive repetition Hidden text with colors, style sheet tricks, etc. Meta-Tags = “… London hotels, hotel, holiday inn, hilton, discount, booking, reservation, sex, mp3, britney spears, viagra, …”
  • 15. Search engine optimization (Spam)  Motives    Operators     Commercial, political, religious, lobbies Promotion funded by advertising budget Contractors (Search Engine Optimizers) for lobbies, companies Web masters Hosting services Forums  E.g., Web master world ( www.webmasterworld.com )   Search engine specific tricks Discussions about academic papers 
  • 17. Features identifying synthetic content  Average word length   Word frequency distribution   Certain words (“the”, “a”, …) appear more often than others N-gram frequency distribution   The mean word length for English prose is about 5 characters; but longer for some forms of keyword stuffing Some words are more likely to occur next to each other than others Grammatical well-formedness  Natural-language parsing is expensive
  • 18. Really good synthetic content “Nigritude Ultramarine”: An SEO competition Links to keep crawlers going Grammatically well-formed but meaningless sentences
  • 19. Content “repurposing”  Content repurposing: The practice of incorporating all or portions of other (unaffiliated) web pages    A “convenient” way to machine generate pages that contain human-authored content Not even necessarily illegal … Two flavors:   Incorporate large portions of a single page Incorporate snippets of multiple pages
  • 20. Example of page-level content “repurposing”
  • 21. Example of phrase-level content “repurposing” How is monetization taking place?
  • 22. Techniques for detecting content repurposing  Single-page flavor: Cluster pages into equivalence classes of very similar pages    If most pages on a site a very similar to pages on other sites, raise a red flag (There are legitimate replicated sites; e.g. mirrors of Linux man pages) Many-snippets flavor: Test if page consists mostly of phrases that also occur somewhere else   Computationally hard problem Probabilistic technique that makes it tractable (Fetterly et al SIGIR 2005 paper)
  • 23. Cloaking   Serve fake content to search engine spider DNS cloaking: Switch IP address. Impersonate Y SPAM Is this a Search Engine spider? Cloaking N Real Doc
  • 24. More spam techniques  Doorway pages   Link spamming     Pages optimized for a single keyword that re-direct to the real target page links between pages that are present for reasons other than merit Mutual admiration societies, hidden links, awards – more on these later Domain flooding: numerous domains that point or re-direct to a target page Robots  Fake query stream – rank checking programs   “Curve-fit” ranking programs of search engines Millions of submissions via Add-Url
  • 25. The war against spam  Quality signals - Prefer authoritative pages based on:  Votes from authors (linkage signals) Votes from users (usage signals)        Anti robot test , i.e. not submitted by a robot Limits on meta-keywords Robust link analysis   Ignore statistically implausible linkage (or text) Use link analysis to detect spammers (guilt by association)   Training set based on known spam Family friendly filters  Policing of URL submissions  Spam recognition by machine learning Linguistic analysis, general classification techniques, etc. For images: flesh tone detectors, source text analysis, etc. Editorial intervention     Blacklists Top queries audited Complaints addressed Suspect pattern detection
  • 26. More on spam  Web search engines have policies on SEO practices they tolerate/block     http://help.yahoo.com/help/us/ysearch/index.html http://www.google.com/intl/en/webmasters/ Adversarial IR: the unending (technical) battle between SEO’s and web search engines Research http://airweb.cse.lehigh.edu/
  • 27. Size of the web
  • 28. What is the size of the web ?  Issues  The web is really infinite      Dynamic content, e.g., calendar Soft 404: www.yahoo.com/<anything> is a valid page Static web contains syntactic duplication, mostly due to mirroring (~30%) Some servers are seldom connected Who cares?    Media, and consequently the user Engine design Engine crawl policy. Impact on recall.
  • 29. What can we attempt to measure? The relative sizes of search engines    The notion of a page being indexed is still reasonably well defined. Already there are problems   Document extension: e.g. engines index pages not yet crawled, by indexing anchortext. Document restriction: All engines restrict what is indexed (first n words, only relevant words, etc.) The coverage of a search engine relative to another particular crawling process. 
  • 30. Graph Structure in the Web http://www9.org/w9cdrom/160/160.html
  • 31. Zipf’s Law on the Web  Number of in-links/out-links to/from a page has a Zipfian distribution.   Length of web pages has a Zipfian distribution.   A large number of docs have the same number of in-links/out-links to/from a page Certain page lengths are most common; i.e. a huge number of docs with the same page length. Number of hits to a web page has a Zipfian distribution.  A large number of pages will generate the same number of hits; dramatic drop-off in number of pages as the above frequency reduces
  • 32. New definition?  (IQ is whatever the IQ tests measure.)   Different engines have different preferences   The statically indexable web is whatever search engines index. max url depth, max count/host, anti-spam rules, priority rules, etc. Different engines index different things under the same URL:  frames, meta-keywords, document restrictions, document extensions, ...
  • 33. Relative Size from Overlap Given two engines A and B Sample URLs randomly from A Check if contained in B and vice versa A∩ B A∩ B = A∩ B = (1/2) * Size A (1/6) * Size B (1/2)*Size A = (1/6)*Size B ∴ Size A / Size B = (1/6)/(1/2) = 1/3 Each test involves: (i) Sampling (ii) Checking
  • 34. Sampling URLs  Ideal strategy: Generate a random URL and check for containment in each index.  Problem: Random URLs are hard to find! Enough to generate a random URL contained in a given Engine. Approach 1: Generate a random URL contained in a given engine     Pick a random query from search log; pick random page returned as search result Suffices for the estimation of relative size Approach 2: Random walks / IP addresses  In theory: might give us a true estimate of the size of the web (as opposed to just relative sizes of indexes)
  • 35. Statistical methods  Approach 1    *** Random queries: preferred method Random searches Approach 2   Random IP addresses Random walks
  • 36. Random URLs from random queries  Generate random query: how?   Not an English Lexicon: 400,000+ words from a web crawl dictionary Conjunctive Queries: w1 and w2 e.g., vocalists AND rsi     Get 100 result URLs from engine A Choose a random URL as the candidate to check for presence in engine B (how do you do this?) This distribution induces a probability weight W(p) for each page. Conjecture: W(SEA) / W(SEB) ~ |SEA| / |SEB|
  • 37. Query Based Checking  Strong Query to check whether an engine B has a document D (corresponding to random URL):     Download D. Get list of words. Use 8 low frequency words as AND query to B Check if D is present in result set. Problems:      Near duplicates (not D, but something identical to D) Frames Redirects Engine time-outs Is 8-word query good enough?
  • 38. Advantages & disadvantages   Statistically sound under the induced weight. Biases induced by random query  Query Bias: Favors content-rich pages in the language(s) of the lexicon    Ranking Bias: Solution: Use conjunctive queries & fetch all Checking Bias: Duplicates, impoverished pages omitted Document or query restriction bias: engine might not deal properly with 8 words conjunctive query   Malicious Bias: Sabotage by engine Operational Problems: Time-outs, failures, engine inconsistencies, index modification.
  • 39. Alt 1: Random searches   These are real queries based on real users, rather than “random queries” Choose random searches extracted from a local log [Lawrence & Giles 97] or build “random searches” [Notess]    Use only queries with small results sets. Count normalized URLs in result sets. Use ratio statistics
  • 40. Advantages & disadvantages  Advantage   Might be a better reflection of the human perception of coverage Issues    Samples are correlated with source of log Duplicates Technical statistical problems (must have nonzero results, ratio average not statistically sound)
  • 41. Random searches    575 & 1050 queries from the NEC RI employee logs 6 Engines in 1998, 11 in 1999 Implementation:  Restricted to queries with < 600 results in total  Counted URLs from each engine after verifying query match  Computed size ratio & overlap for individual queries  Estimated index size ratio & overlap by averaging over all queries
  • 42. Queries from Lawrence and Giles study           adaptive access control neighborhood preservation topographic hamiltonian structures right linear grammar pulse width modulation neural unbalanced prior probabilities ranked assignment method internet explorer favourites importing karvel thornber zili liu           softmax activation function bose multidimensional system theory gamma mlp dvi2pdf john oliensis rieke spikes exploring neural video watermarking counterpropagation network fat shattering dimension abelson amorphous computing
  • 43. Alt 2: Random IP addresses   Generate random IP addresses Find a web server at the given address   If there’s one Collect all pages from server  From this, choose a page at random
  • 44. Random IP addresses  HTTP requests to random IP addresses    Ignored: empty or authorization required or excluded [Lawr99] Estimated 2.8 million IP addresses running crawlable web servers (16 million total) from observing 2500 servers. OCLC using IP sampling found 8.7 M hosts in 2001   Netcraft [Netc02] accessed 37.2 million hosts in July 2002 [Lawr99] exhaustively crawled 2500 servers and extrapolated   Estimated size of the web to be 800 million Estimated use of metadata descriptors:  Meta tags (keywords, description) in 34% of home pages, Dublin core metadata in 0.3%
  • 45. Advantages & disadvantages  Advantages    Clean statistics Independent of crawling strategies Disadvantages    Doesn’t deal with duplication Many hosts might share one IP, or not accept requests No guarantee all pages are linked to root page.   Power law for # pages/hosts generates bias towards sites with few pages.   Eg: employee pages But bias can be accurately quantified IF underlying distribution understood Potentially influenced by spamming (multiple IP’s for same server to avoid IP block)
  • 46. Alt 3: Random walks   View the Web as a directed graph Build a random walk on this graph  Includes various “jump” rules back to visited sites    Converges to a stationary distribution      Does not get stuck in spider traps! Can follow all links! Must assume graph is finite and independent of the walk. Conditions are not satisfied (cookie crumbs, flooding) Time to convergence not really known Sample from stationary distribution of walk Use the “strong query” method to check coverage by SE
  • 47. Advantages & disadvantages  Advantages    “Statistically clean” method at least in theory! Could work even for infinite web (assuming convergence) under certain metrics. Disadvantages    List of seeds is a problem. Practical approximation might not be valid. Non-uniform distribution  Subject to link spamming
  • 48. Conclusions     No sampling solution is perfect. Lots of new ideas ... ....but the problem is getting harder Quantitative studies are fascinating and a good research problem
  • 50. Duplicate documents   The web is full of duplicated content Strict duplicate detection = exact match   Not as common But many, many cases of near duplicates  E.g., Last modified date the only difference between two copies of a page
  • 51. Duplicate/Near-Duplicate Detection   Duplication: Exact match can be detected with fingerprints (64-bit), if fingerprints match, check further Near-Duplication: Approximate match  Overview   Compute syntactic similarity with an edit-distance measure Use similarity threshold to detect near-duplicates   E.g., Similarity > 80% => Documents are “near duplicates” Not transitive though sometimes used transitively
  • 52. Computing Similarity   Features:  Segments of a document (natural or artificial breakpoints)  Shingles (Word N-Grams)  a rose is a rose is a rose → a_rose_is_a rose_is_a_rose is_a_rose_is a_rose_is_a Similarity Measure between two docs (= sets of shingles)  Set intersection  Specifically (Size_of_Intersection / Size_of_Union)
  • 53. Shingles + Set Intersection Computing exact set intersection of shingles between all pairs of documents is expensive/intractable   Approximate using a cleverly chosen subset of shingles from each (a sketch) Estimate (size_of_intersection / size_of_union) based on a short sketch  Doc Doc A A Shingle set A Doc Doc B B Shingle set A Sketch A Jaccard Sketch A
  • 54. Sketch of a document  Create a “sketch vector” (of size ~200) for each document   Documents that share ≥ t (say 80%) corresponding vector elements are near duplicates For doc D, sketchD[ i ] is as follows:    Let f map all shingles in the universe to 0..2m (e.g., f = fingerprinting) Let πi be a random permutation on 0..2m Pick MIN {πi(f(s))} over all shingles s in D
  • 55. Computing Sketch[i] for Doc1 Document 1 264 Start with 64-bit f(shingles) 264 Permute on the number line 2 64 264 with πi Pick the min value
  • 56. Test if Doc1.Sketch[i] = Doc2.Sketch[i] Document 2 Document 1 264 264 264 264 A 264 264 2 64 B Are these equal? Test for 200 random permutations: π1, π2,… π200 264
  • 57. However… Document 2 Document 1 264 264 264 264 A 264 264 B 264 264 A = B iff the shingle with the MIN value in the union of Doc1 and Doc2 is common to both (i.e., lies in the intersection) Why? Claim: This happens with probability Size_of_intersection / Size_of_union
  • 58. Set Similarity of sets Ci , Cj Jaccard(Ci , C j ) =   Ci  C j Ci  C j View sets as columns of a matrix A; one row for each element in the universe. aij = 1 indicates presence of item i in set j C1 C2 Example 0 1 1 0 1 1 0 1 0 1 Jaccard(C 1 ,C 2 ) = 2/5 = 0.4
  • 59. Key Observation  For columns Ci, Cj, four types of rows Ci A B C D   1 1 0 0 Cj 1 0 1 0 Overload notation: A = # of rows of type A Claim A Jaccard(Ci , C j ) = A+B+C
  • 60. “Min” Hashing   Randomly permute rows Hash h(Ci) = index of first row with 1 in column Ci  Surprising Property  Why? P [ h(Ci ) = h(C j ) ] = Jaccard ( Ci , C j )    Both are A/(A+B+C) Min: Look down columns Ci, Cj until first non-Type-D row h(Ci) = h(Cj)  type A row  Because we chose a random permutation, prob of this occurring is A/(A+B+C)
  • 61. Min-Hash sketches   Pick P random row permutations MinHash sketch SketchD = list of P (usually, 200) indexes (i.e. integers) of first rows with 1 in column C  Similarity of signatures   Let sim[sketch(Ci),sketch(Cj)] = fraction of permutations where MinHash values agree Observe E[sim(sig(Ci),sig(Cj))] = Jaccard(Ci,Cj)
  • 62. Example R1 R2 R3 R4 R5 C1 1 0 1 1 0 C2 0 1 0 0 1 C3 1 1 0 1 0 Signatures (composed of min) S1 S2 S3 Perm 1 = (12345) 1 2 1 Perm 2 = (54321) 4 5 4 Perm 3 = (34512) 3 5 4 Similarities 1-2 1-3 2-3 Col-Col 0.00 0.50 0.25 Sig-Sig 0.00 0.67 0.00 Good estimate
  • 63. Implementation Trick   Permuting universe even once is prohibitive Row Hashing    Pick P hash functions hk: {1,…,n}{1,…,O(n)} Ordering under hk gives random permutation of rows One-pass Implementation  For each Ci and hk, keep “slot” for min-hash value  Initialize all slot(Ci,hk) to infinity  Scan rows in arbitrary order looking for 1’s  Suppose row Rj has 1 in column Ci  For each hk,
  • 64. Example R1 R2 R3 R4 R5 C1 1 0 1 1 0 C2 0 1 1 0 1 h(x) = x mod 5 g(x) = 2x+1 mod 5 h(1) = 1 g(1) = 3 C1 slots C2 slots 1 3 - h(2) = 2 g(2) = 0 1 3 2 0 h(3) = 3 g(3) = 2 1 2 2 0 h(4) = 4 g(4) = 4 h(5) = 0 g(5) = 1 1 2 1 2 2 0 0 0
  • 65. Comparing Signatures   Signature Matrix S  Rows = Hash Functions  Columns = Columns  Entries = Signatures Can compute – Pair-wise similarity of any pair of signature columns
  • 66. All signature pairs   Now we have an extremely efficient method for estimating a Jaccard coefficient for a single pair of documents. But we still have to estimate N2 coefficients where N is the number of web pages.    Still slow One solution: locality sensitive hashing (LSH) Another solution: sorting (Henzinger 2006)

Editor's Notes

  • #19: Wikipedia has a good description of the Nigritude Ultramarine contest, which ran from 5/7/2004 to 7/7/2004 and drew 200 contestants.
  • #22: Unfortunately, I can’t show you just how the first replicating site monetizes on the Paris Hilton sentences, but you might be able to figure it out yourselves. The second site is even more amazing: Not only does it try to sell vacations to Paris by invoking Paris Hilton’s name, but it also manages to put Miss Hilton and Condoleezza Rice into the same sentence!
  • #59: Union = 7-2 = 5. The number of unique set items in both sets