Εύρεση & ιαχείριση Πληροφορίας στον Παγκόσµιο Ιστό ιδάσκων ηµήτριος Κατσαρός, Ph.D. @ Τµ. Μηχανικών Η/Υ, Τηλεπικοινωνιών & ικτύων Πανεπιστήµιο Θεσσαλίας ιάλεξη 7η: 17/04/2007 1 Web characteristics 2 Search use (iprospect Survey, 4/04, http://www.iprospect.com/premiumpdfs/iprospectsurveycomplete.pdf) 3 1
Without search engines the web wouldn t scale 1. No incentive in creating content unless it can be easily found other finding methods haven t kept pace (taxonomies, bookmarks, etc) 2. The web is both a technology artifact and a social environment The Web has become the new normal in the American way of life; those who don t go online constitute an ever-shrinking minority. [Pew Foundation report, January 2005] 3. Search engines make aggregation of interest possible: Create incentives for very specialized niche players Economical specialized stores, providers, etc Social narrow interests, specialized communities, etc 4. The acceptance of search interaction makes unlimited selection stores possible: Amazon, Netflix, etc 5. Search turned out to be the best mechanism for advertising on the web, a $15+ B industry. Growing very fast but entire US advertising industry $250B huge room to grow Sponsored search marketing is about $10B 4 Classical IR vs. Web IR 5 Basic assumptions of Classical Information Retrieval Corpus: Fixed document collection Goal: Retrieve documents with information content that is relevant to user s information need 6 2
Classic IR Goal Classic relevance For each query Q and stored document D in a given corpus assume there exists relevance Score(Q, D) Score is average over users U and contexts C Optimize Score(Q, D) as opposed to Score(Q, D, U, C) That is, usually: Context ignored Individuals ignored Corpus predetermined Bad assumptions in the web context 7 Web IR 8 The coarse-level dynamics Advertisement Editorial Feeds Crawls Subscription Transaction Content creators Content aggregators Content consumers 9 3
Brief (non-technical) history Early keyword-based engines Altavista, Excite, Infoseek, Inktomi, ca. 1995-1997 Paid placement ranking: Goto.com (morphed into Overture.com Yahoo!) Your search ranking depended on how much you paid Auction for keywords: casino was expensive! 10 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 follows suit, acquiring Overture (for paid placement) and Inktomi (for search) 11 Ads Algorithmic results. 12 4
Sponsored Links CG Appliance Express Discount Appliances (650) 756-3931 Same Day Certified Installation www.cgappliance.com San Francisco-Oakland-San Jose, CA 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 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 Ads vs. search results Google has maintained that ads (based on vendors bidding for keywords) do not affect vendors rankings in search results Sponsored Links CG Appliance Express Discount Appliances (650) 756-3931 Same Day Certified Installation www.cgappliance.com San Francisco-Oakland-San Jose, CA 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) Search = miele Miele, Inc -- Anything else is a compromise 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 13 Ads vs. search results Other vendors (Yahoo, MSN) have made similar statements from time to time Any of them can change anytime We will focus primarily on search results independent of paid placement ads Although the latter is a fascinating technical subject in itself 14 Web search basics User Web spider Search Indexer The Web Indexes Ad indexes 15 5
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 Downloads Shop Seattle weather Mars surface images Gray areas Find a good hub Canon S410 Exploratory search see what s there Car rental Brasil 16 Web search users Make ill defined queries Short AV 2001: 2.54 terms avg, 80% < 3 words) AV 1998: 2.35 terms avg, 88% < 3 words [Silv98] Imprecise terms Sub-optimal syntax (most queries without operator) Low effort Wide variance in Needs Expectations Knowledge Bandwidth Specific behavior 85% look over one result screen only (mostly above the fold) 78% of queries are not modified (one query/session) Follow links the scent of information... 17 Query Distribution Power law: few popular broad queries, many rare specific queries 18 6
How far do people look for results? (Source: iprospect.com WhitePaper_2006_SearchEngineUserBehavior.pdf) 19 True example* Mis-translation TASK Info Need Verbal form Query Noisy building fan in courtyard Mis-conception Info about EPA regulations Mis-formulation What are the EPA rules about noise pollution EPA sound pollution Query Refinement SEARCH ENGINE Results Corpus Polysemy Synonimy * To Google or to GOTO, Business Τµ. Week Μηχανικών Online, Η/Υ, September Τηλεπικοινωνιών 28, & 2001 ικτύων, Πανεπιστήµιο Θεσσαλίας 20 Users empirical evaluation of results Quality of pages varies widely Relevance is not enough Other desirable qualities (non IR!!) Content: Trustworthy, new info, non-duplicates, well maintained, Web readability: display correctly & fast No annoyances: pop-ups, etc 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 Recall matters when the number of matches is very small User perceptions may be unscientific, but are significant over a large aggregate 21 7
Users empirical evaluation of engines Relevance and validity of results UI Simple, no clutter, error tolerant Trust Results are objective Coverage of topics for poly-semic queries Pre/Post process tools provided Mitigate user errors (auto spell check, syntax errors, ) 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 22 Loyalty to a given search engine (iprospect Survey, 4/04) 23 The Web corpus 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 corpora but corporate records are catching up. Growth slowed down from initial volume doubling every few months but still expanding Content can be dynamically generated 24 8
The Web: Dynamic content A page without a static html version E.g., current status of flight AA129 Current availability of rooms at a hotel Usually, assembled at the time of a request from a browser Typically, URL has a? character in it AA129 Application server Browser Back-end databases 25 Dynamic content Most dynamic content is ignored by web spiders Many reasons including malicious spider traps Some dynamic content (news stories from subscriptions) are sometimes delivered as dynamic content Application-specific spidering Spiders commonly view web pages just as Lynx (a text browser) would Note: even static pages are typically assembled on the fly (e.g., headers are common) 26 The web: size What is being measured? Number of hosts Number of (static) html pages Volume of data Number of hosts netcraft survey http://news.netcraft.com/archives/web_server_survey.html Monthly report on how many web hosts & servers are out there Number of pages numerous estimates (will discuss later) 27 9
Netcraft Web Server Survey http://news.netcraft.com/archives/web_server_survey.html 28 The web: evolution All of these numbers keep changing Relatively few scientific studies of the evolution of the web [Fetterly & al, 2003] http://research.microsoft.com/research/sv/sv-pubs/p97- fetterly/p97-fetterly.pdf Sometimes possible to extrapolate from small samples (fractal models) [Dill & al, 2001] http://www.vldb.org/conf/2001/p069.pdf 29 Rate of change [Cho00] 720K pages from 270 popular sites sampled daily from Feb 17 Jun 14, 1999 Any changes: 40% weekly, 23% daily [Fett02] Massive study 151M pages checked over few months Significant changed -- 7% weekly Small changes 25% weekly [Ntul04] 154 large sites re-crawled from scratch weekly 8% new pages/week 8% die 5% new content 25% new links/week 30 10
Static pages: rate of change Fetterly et al. study (2002): several views of data, 150 million pages over 11 weekly crawls Bucketed into 85 groups by extent of change 31 Other characteristics Significant duplication Syntactic 30%-40% (near) duplicates [Brod97, Shiv99b, etc.] Semantic??? High linkage More than 8 links/page in the average Complex graph topology Not a small world; bow-tie structure [Brod00] Spam Billions of pages 32 Spam Search Engine Optimization 33 11
The trouble with paid placement It costs money. What s the alternative? Search Engine Optimization: Tuning your web page to rank highly in the 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 34 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 35 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, 36 12
Search engine optimization (Spam) Motives Commercial, political, religious, lobbies Promotion funded by advertising budget Operators 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 37 Cloaking Serve fake content to search engine spider DNS cloaking: Switch IP address. Impersonate Y SPAM Cloaking Is this a Search Engine spider? N Real Doc 38 The spam industry 39 13
40 More spam techniques Doorway pages Pages optimized for a single keyword that re-direct to the real target page Link spamming 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 41 The war against spam Quality signals - Prefer authoritative pages based on: Votes from authors (linkage signals) Votes from users (usage signals) Policing of URL submissions Anti robot test Limits on meta-keywords Robust link analysis Ignore statistically implausible linkage (or text) Use link analysis to detect spammers (guilt by association) Spam recognition by machine learning Training set based on known spam Family friendly filters 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 42 14
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/ 43 Answering the need behind the query Semantic analysis Query language determination Auto filtering Different ranking (if query in Japanese do not return English) Hard & soft (partial) matches Personalities (triggered on names) Cities (travel info, maps) Medical info (triggered on names and/or results) Stock quotes, news (triggered on stock symbol) Company info Etc. Natural Language reformulation Integration of Search and Text Analysis 44 The spatial context -- geo-search Two aspects Geo-coding -- encode geographic coordinates to make search effective Geo-parsing -- the process of identifying geographic context. Geo-coding Geometrical hierarchy (squares) Natural hierarchy (country, state, county, city, zip-codes, etc) Geo-parsing Pages (infer from phone nos, zip, etc). About 10% can be parsed. Queries (use dictionary of place names) Users Explicit (tell me your location -- used by NL, registration, from ISP) From IP data Mobile phones In its infancy, many issues (display size, privacy, etc) 45 15
Yahoo!: britney spears 46 Ask Jeeves: las vegas 47 Yahoo!: salvador hotels 48 16
Yahoo shortcuts Various types of queries that are understood 49 Google andrei broder new york 50 Answering the need behind the query : Context Context determination spatial (user location/target location) query stream (previous queries) personal (user profile) explicit (user choice of a vertical search, ) implicit (use Google from France, use google.fr) Context use Result restriction Kill inappropriate results Ranking modulation Use a rough generic ranking, but personalize later 51 17
Google: dentists bronx 52 Yahoo!: dentists (bronx) 53 54 18
Query expansion 55 Context transfer 56 No transfer 57 19
Context transfer 58 Transfer from search results 59 60 20
Resources IIR Chapter 19 61 Web characteristics II: Web size measurement Near-duplicate detection 62 Today s topics Estimating web size and search engine index size Near-duplicate document detection 63 21
Size of the web 64 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. 65 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. 66 22
New definition? (IQ is whatever the IQ tests measure.) The statically indexable web is whatever search engines index. Different engines have different preferences 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,... 67 Statistical methods Random queries Random searches Random IP addresses Random walks 68 Relative Size from Overlap [Bharat & Broder, 98] Sample URLs randomly from A Check if contained in B and vice versa A B A B = (1/2) * Size A A B = (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 69 23
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. Key lesson from this lecture. 70 Random URLs from random queries [Bharat & B, 98] Generate random query: how? Lexicon: 400,000+ words from a crawl of Yahoo! Conjunctive Queries: w 1 and w 2 e.g., vocalists AND rsi Get 100 result URLs from the source engine Choose a random URL as the candidate to check for presence in other engines. This distribution induces a probability weight W(p) for each page. Conjecture: W(SE 1 ) / W(SE 2 ) ~ SE 1 / SE 2 71 Query Based Checking Strong Query to check for a document D: Download document. Get list of words. Use 8 low frequency words as AND query Check if D is present in result set. Problems: Near duplicates Frames Redirects Engine time-outs Might be better to use e.g. 5 distinct conjunctive queries of 6 words each. 72 24
Computing Relative Sizes and Total Coverage [BB98] a = AltaVista, e = Excite, h = HotBot, i = Infoseek f xy = fraction of x in y Six pair-wise overlaps f ah * a - f ha * h = ε 1 f ai * a - f ia * i = ε 2 f ae * a - f ea * e = ε 3 f hi * h - f ih * i = ε 4 f he * h - f eh * e = ε 5 f ei * e - f ie * i = ε 6 Arbitrarily, let a = 1. We have 6 equations and 3 unknowns. Solve for e, h and i to minimize Σ ε i 2 Compute engine overlaps. Re-normalize so that the total joint coverage is 100% 73 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. 74 Random searches 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 75 25
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 non-zero results, ratio average, use harmonic mean?) 76 Random searches [Lawr98, Lawr99] 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 77 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 78 26
Random IP addresses [Lawrence & Giles 99] Generate random IP addresses Find a web server at the given address If there s one Collect all pages from server. Method first used by O Neill, McClain, & Lavoie, A Methodology for Sampling the World Wide Web, 1997. http://digitalarchive.oclc.org/da/viewobject.jsp?objid=0000003 447 79 Random IP addresses [ONei97, Lawr99] 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% 80 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. Eg: employee pages Power law for # pages/hosts generates bias towards sites with few 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) 81 27
Random walks [Henzinger et al WWW9] View the Web as a directed graph Build a random walk on this graph Includes various jump rules back to visited sites Does not get stuck in spider traps! Can follow all links! Converges to a stationary distribution 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 82 Dependence on seed list How well connected is the graph? [Broder et al., WWW9] 83 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 84 28
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 85 Duplicate detection 86 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 87 29
Duplicate/Near-Duplicate Detection Duplication: Exact match can be detected with fingerprints 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 88 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 Similarity Measure between two docs (= sets of shingles) Set intersection [Brod98] (Specifically, Size_of_Intersection / Size_of_Union ) Jaccard measure 89 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 90 30
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, sketch D [ i ] is as follows: Let f map all shingles in the universe to 0..2 m (e.g., f = fingerprinting) Let π i be a random permutation on 0..2 m Pick MIN {π i (f(s))} over all shingles s in D 91 Computing Sketch[i] for Doc1 Document 1 2 64 2 64 2 64 Start with 64-bit f(shingles) Permute on the number line with π i 2 64 Pick the min value 92 Test if Doc1.Sketch[i] = Doc2.Sketch[i] Document 1 Document 2 2 64 2 64 2 64 2 64 2 64 2 64 A 2 64 B 2 64 Are these equal? Test for 200 random permutations: π 1, π 2, π 200 93 31
However Document 1 Document 2 2 64 2 64 2 64 2 64 A 2 64 2 64 B 2 64 2 64 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) This happens with probability: Size_of_intersection / Size_of_union Why? 94 Set Similarity Set Similarity (Jaccard measure) Ci I C j Jaccard(C i, C j) = Ci U C j View sets as columns of a matrix; one row for each element in the universe. a ij = 1 indicates presence of item i in set j Example C 1 C 2 0 1 1 0 1 1 Jaccard(C 1,C 2 ) = 2/5 = 0.4 0 0 1 1 0 1 95 Key Observation For columns C i, C j, four types of rows C i C j A 1 1 B 1 0 C 0 1 D 0 0 Overload notation: A = # of rows of type A Claim A Jaccard(Ci,C j) = A + B + C 96 32
Min Hashing Randomly permute rows Hash h(c i ) = index of first row with 1 in column C i Surprising Property P [ h(c i ) = h(c j ) ] = Jaccard( C i, C j ) Why? Both are A/(A+B+C) Look down columns C i, C j until first non-type-d row h(c i ) = h(c j ) type A row 97 Min-Hash sketches Pick P random row permutations MinHash sketch Sketch D = list of P indexes of first rows with 1 in column C Similarity of signatures Let sim[sketch(c i ),sketch(c j )] = fraction of permutations where MinHash values agree Observe E[sim(sig(C i ),sig(c j ))] = Jaccard(C i,c j ) 98 Example C 1 C 2 C 3 R 1 1 0 1 R 2 0 1 1 R 3 1 0 0 R 4 1 0 1 R 5 0 1 0 Signatures S 1 S 2 S 3 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 99 33
Implementation Trick Permuting rows even once is prohibitive Row Hashing Pick P hash functions h k : {1,,n} {1,,O(n)} Ordering under h k gives random row permutation One-pass Implementation For each C i and h k, keep slot for min-hash value Initialize all slot(c i,h k ) to infinity Scan rows in arbitrary order looking for 1 s Suppose row R j has 1 in column C i For each h k, if h k (j) < slot(c i,h k ), then slot(c i,h k ) h k (j) 100 Example C 1 C 2 R 1 1 0 R 2 0 1 R 3 1 1 R 4 1 0 R 5 0 1 h(x) = x mod 5 g(x) = 2x+1 mod 5 C 1 slots C 2 slots h(1) = 1 1 - g(1) = 3 3 - h(2) = 2 1 2 g(2) = 0 3 0 h(3) = 3 1 2 g(3) = 2 2 0 h(4) = 4 1 2 g(4) = 4 2 0 h(5) = 0 1 0 g(5) = 1 2 0 101 Comparing Signatures Signature Matrix S Rows = Hash Functions Columns = Columns Entries = Signatures Compute Pair-wise similarity of signature columns Problem MinHash fits column signatures in memory But comparing signature-pairs takes too much time Technique to limit candidate pairs? Locality Sensitive Hashing (LSH) 102 34
Resources IIR 19 See also Phelps & Wilensky. Robust Hyperlinks & Locations, 2002 Ziv Bar-Yossef and Maxim Gurevich. Random Sampling from a Search Engine s Index, WWW 2006. Broder et al. Estimating corpus size via queries. CIKM 2006. 103 More resources Related papers: [Bar Yossef & al, VLDB 2000], [Rusmevichientong & al, 2001], [Bar Yossef & al, 2003] 104 Crawling and web indexes 105 35
Today s lecture Crawling Connectivity servers 106 Basic crawler operation Begin with known seed pages Fetch and parse them Extract URLs they point to Place the extracted URLs on a queue Fetch each URL on the queue and repeat 107 Crawling picture URLs crawled and parsed Unseen Web Web Seed pages URLs frontier 108 36
Simple picture complications Web crawling isn t feasible with one machine All of the above steps distributed Even non-malicious pages pose challenges Latency/bandwidth to remote servers vary Webmasters stipulations How deep should you crawl a site s URL hierarchy? Site mirrors and duplicate pages Malicious pages Spam pages Spider traps incl dynamically generated Politeness don t hit a server too often 109 What any crawler must do Be Polite: Respect implicit and explicit politeness considerations for a website Only crawl pages you re allowed to Respect robots.txt (more on this shortly) Be Robust: Be immune to spider traps and other malicious behavior from web servers 110 What any crawler should do Be capable of distributed operation: designed to run on multiple distributed machines Be scalable: designed to increase the crawl rate by adding more machines Performance/efficiency: permit full use of available processing and network resources 111 37
What any crawler should do Fetch pages of higher quality first Continuous operation: Continue fetching fresh copies of a previously fetched page Extensible: Adapt to new data formats, protocols 112 Updated crawling picture URLs crawled and parsed Unseen Web Seed Pages URL frontier Crawling thread 113 URL frontier Can include multiple pages from the same host Must avoid trying to fetch them all at the same time Must try to keep all crawling threads busy 114 38
Explicit and implicit politeness Explicit politeness: specifications from webmasters on what portions of site can be crawled robots.txt Implicit politeness: even with no specification, avoid hitting any site too often 115 Robots.txt Protocol for giving spiders ( robots ) limited access to a website, originally from 1994 www.robotstxt.org/wc/norobots.html Website announces its request on what can(not) be crawled For a URL, create a file URL/robots.txt This file specifies access restrictions 116 Robots.txt example No robot should visit any URL starting with "/yoursite/temp/", except the robot called searchengine": User-agent: * Disallow: /yoursite/temp/ User-agent: searchengine Disallow: 117 39
Processing steps in crawling Pick a URL from the frontier Fetch the document at the URL Parse the URL Extract links from it to other docs (URLs) Check if URL has content already seen If not, add to indexes E.g., only crawl.edu, For each extracted URL obey robots.txt, etc. Ensure it passes certain URL filter tests Check if it is already in the frontier (duplicate URL elimination) 118 Basic crawl architecture DNS Doc FP s robots filters URL set WWW Fetch Parse Content seen? URL filter Dup URL elim URL Frontier 119 DNS (Domain Name Server) A lookup service on the internet Given a URL, retrieve its IP address Service provided by a distributed set of servers thus, lookup latencies can be high (even seconds) Common OS implementations of DNS lookup are blocking: only one outstanding request at a time Solutions DNS caching Batch DNS resolver collects requests and sends them out together 120 40
Parsing: URL normalization When a fetched document is parsed, some of the extracted links are relative URLs E.g., at http://en.wikipedia.org/wiki/main_page we have a relative link to /wiki/wikipedia:general_disclaimer which is the same as the absolute URL http://en.wikipedia.org/wiki/wikipedia:general_disclaimer During parsing, must normalize (expand) such relative URLs 121 Content seen? Duplication is widespread on the web If the page just fetched is already in the index, do not further process it This is verified using document fingerprints or shingles 122 Filters and robots.txt Filters regular expressions for URL s to be crawled/not Once a robots.txt file is fetched from a site, need not fetch it repeatedly Doing so burns bandwidth, hits web server Cache robots.txt files 123 41
Duplicate URL elimination For a non-continuous (one-shot) crawl, test to see if an extracted+filtered URL has already been passed to the frontier For a continuous crawl see details of frontier implementation 124 Distributing the crawler Run multiple crawl threads, under different processes potentially at different nodes Geographically distributed nodes Partition hosts being crawled into nodes Hash used for partition How do these nodes communicate? 125 Communication between nodes The output of the URL filter at each node is sent to the Duplicate URL Eliminator at all nodes DNS Doc FP s robots filters To other hosts URL set WWW Fetch Parse Content seen? URL Frontier URL filter Host splitter From other hosts Dup URL elim 126 42
URL frontier: two main considerations Politeness: do not hit a web server too frequently Freshness: crawl some pages more often than others E.g., pages (such as News sites) whose content changes often These goals may conflict each other. (E.g., simple priority queue fails many links out of a page go to its own site, creating a burst of accesses to that site.) 127 Politeness challenges Even if we restrict only one thread to fetch from a host, can hit it repeatedly Common heuristic: insert time gap between successive requests to a host that is >> time for most recent fetch from that host 128 URL frontier: Mercator scheme URLs Prioritizer K front queues Biased front queue selector Back queue router B back queues Single host on each Back queue selector Crawl thread requesting URL 129 43
Mercator URL frontier URL s flow in from the top into the frontier Front queues manage prioritization Back queues enforce politeness Each queue is FIFO 130 Front queues Prioritizer 1 K Biased front queue selector Back queue router 131 Front queues Prioritizer assigns to URL an integer priority between 1 and K Appends URL to corresponding queue Heuristics for assigning priority Refresh rate sampled from previous crawls Application-specific (e.g., crawl news sites more often ) 132 44
Biased front queue selector When a back queue requests a URL (in a sequence to be described): picks a front queue from which to pull a URL This choice can be round robin biased to queues of higher priority, or some more sophisticated variant Can be randomized 133 Back queues Biased front queue selector Back queue router 1 B Back queue selector Heap 134 Back queue invariants Each back queue is kept non-empty while the crawl is in progress Each back queue only contains URLs from a single host Maintain a table from hosts to back queues Host name Back queue 3 1 B 135 45
Back queue heap One entry for each back queue The entry is the earliest time t e at which the host corresponding to the back queue can be hit again This earliest time is determined from Last access to that host Any time buffer heuristic we choose 136 Back queue processing A crawler thread seeking a URL to crawl: Extracts the root of the heap Fetches URL at head of corresponding back queue q (look up from table) Checks if queue q is now empty if so, pulls a URL v from front queues If there s already a back queue for v s host, append v to q and pull another URL from front queues, repeat Else add v to q When q is non-empty, create heap entry for it 137 Number of back queues B Keep all threads busy while respecting politeness Mercator recommendation: three times as many back queues as crawler threads 138 46
Connectivity servers 139 Connectivity Server [CS1: Bhar98b, CS2 & 3: Rand01] Support for fast queries on the web graph Which URLs point to a given URL? Which URLs does a given URL point to? Stores mappings in memory from URL to outlinks, URL to inlinks Applications Crawl control Web graph analysis Connectivity, crawl optimization Link analysis 140 Most recent published work Boldi and Vigna http://www2004.org/proceedings/docs/1p595.pdf Webgraph set of algorithms and a java implementation Fundamental goal maintain node adjacency lists in memory For this, compressing the adjacency lists is the critical component 141 47
Adjacency lists The set of neighbors of a node Assume each URL represented by an integer E.g., for a 4 billion page web, need 32 bits per node Naively, this demands 64 bits to represent each hyperlink 142 Adjaceny list compression Properties exploited in compression: Similarity (between lists) Locality (many links from a page go to nearby pages) Use gap encodings in sorted lists Distribution of gap values 143 Storage Boldi/Vigna get down to an average of ~3 bits/link (URL to URL edge) For a 118M node web graph 144 48
Main ideas of Boldi/Vigna Consider lexicographically ordered list of all URLs, e.g., www.stanford.edu/alchemy www.stanford.edu/biology www.stanford.edu/biology/plant www.stanford.edu/biology/plant/copyright www.stanford.edu/biology/plant/people www.stanford.edu/chemistry 145 Boldi/Vigna Each of these URLs has an adjacency list Main thesis: because of templates, the adjacency list of a node is similar to one of the 7 preceding URLs in the lexicographic ordering Express adjacency list in terms of one of these E.g., consider these adjacency lists 1, 2, 4, 8, 16, 32, 64 1, 4, 9, 16, 25, 36, 49, 64 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144 1, 4, 8, 16, 25, 36, 49, 64 Encode as (-2), remove 9, add 8 146 Resources IIR Chapter 20 http://research.microsoft.com/~najork/mercator.pdf www.robotstxt.org/wc/norobots.html http://www.www2004.org/proceedings/docs/1p595.pdf 147 49