Question? Leave a message!




Retrieval Search engine optimization

Retrieval Search engine optimization
Introduction to Information Retrieval Introduction to Information Retrieval Web search basics www.ThesisScientist.comIntroduction to Information Retrieval Brief (nontechnical) history  Early keywordbased engines ca. 19951997  Altavista, Excite, Infoseek, Inktomi, Lycos  Paid search ranking: Goto (morphed into Overture.com  Yahoo)  Your search ranking depended on how much you paid  Auction for keywords: casino was expensive www.ThesisScientist.comIntroduction to Information Retrieval Brief (nontechnical) history  1998+: Linkbased 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 search “ads” to the side, independent of search results  Yahoo followed suit, acquiring Overture (for paid placement) and Inktomi (for search)  2005+: Google gains search share, dominating in Europe and very strong in North America  2009: Yahoo and Microsoft propose combined paid search offering www.ThesisScientist.comIntroduction to Information Retrieval Paid Search Ads Algorithmic results. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.4.1 Web search basics Sponsored Links CG Appliance Express Discount Appliances (650) 7563931 Same Day Certified Installation www.cgappliance.com San FranciscoOaklandSan Jose, CA User Miele Vacuum Cleaners Miele Vacuums Complete Selection Free Shipping www.vacuums.com Miele Vacuum Cleaners MieleFree Air shipping All models. Helpful advice. www.bestvacuum.com Results 1 10 of about 7,310,000 for miele. (0.12 seconds) Web 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 Web spider Miele Deutscher Hersteller von Einbaugeräten, Hausgeräten ... Translate this page Das Portal zum Thema Essen Geniessen online unter www.zutisch.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 www.ThesisScientist.com Indexes Ad indexesIntroduction to Information Retrieval Sec. 19.4.1 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 (webmediated) (35 / 20)  Access a service Seattle weather Mars surface images  Downloads Canon S410  Shop  Gray areas Car rental Brasil  Find a good hub  Exploratory search “see what’s there” www.ThesisScientist.comIntroduction to Information Retrieval How far do people look for results (Source: iprospect.com WhitePaper2006SearchEngineUserBehavior.pdf) www.ThesisScientist.comIntroduction to Information Retrieval Users’ empirical evaluation of results  Quality of pages varies widely  Relevance is not enough  Other desirable qualities (non IR)  Content: Trustworthy, diverse, nonduplicated, well maintained  Web readability: display correctly fast  No annoyances: popups, 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 www.ThesisScientist.comIntroduction to Information Retrieval 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, spellcheck, etc.  Web addresses typed in the search box  “The first, the last, the best and the worst …” www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2 The Web document collection  No design/coordination  Distributed content creation, linking, democratization of publishing  Content includes truth, lies, obsolete information, contradictions …  Unstructured (text, html, …), semi structured (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 The Web still expanding  Content can be dynamically generated www.ThesisScientist.comIntroduction to Information Retrieval SPAM (SEARCH ENGINE OPTIMIZATION) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 The trouble with paid search ads …  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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 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  www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 Simplest forms  First generation engines relied heavily on tf/idf  The topranked 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., mauiresort 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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 Variants of keyword stuffing  Misleading metatags, excessive repetition  Hidden text with colors, style sheet tricks, etc. MetaTags = “… London hotels, hotel, holiday inn, hilton, discount, booking, reservation, sex, mp3, britney spears, viagra, …” www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 Cloaking  Serve fake content to search engine spider  DNS cloaking: Switch IP address. Impersonate SPAM N Is this a Search Engine spider Real Y Cloaking Doc www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.2.2 More spam techniques  Doorway pages  Pages optimized for a single keyword that redirect 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  “Curvefit” ranking programs of search engines  Millions of submissions via AddUrl www.ThesisScientist.comIntroduction to Information Retrieval The war against spam  Quality signals Prefer  Spam recognition by authoritative pages based machine learning  Training set based on known on: spam  Votes from authors (linkage  Family friendly filters signals)  Linguistic analysis, general  Votes from users (usage signals) classification techniques, etc.  Policing of URL submissions  For images: flesh tone detectors, source text analysis,  Anti robot test etc.  Limits on metakeywords  Editorial intervention  Robust link analysis  Blacklists  Ignore statistically implausible  Top queries audited linkage (or text)  Complaints addressed  Use link analysis to detect  Suspect pattern detection spammers (guilt by association) www.ThesisScientist.comIntroduction to Information Retrieval 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/ www.ThesisScientist.comIntroduction to Information Retrieval SIZE OF THE WEB www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 What is the size of the web  Issues  The web is really infinite  Dynamic content, e.g., calendars  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. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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.) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 New definition  The statically indexable web is whatever search engines index.  IQ is whatever the IQ tests measure.  Different engines have different preferences  max url depth, max count/host, antispam rules, priority rules, etc.  Different engines index different things under the same URL:  frames, metakeywords, document restrictions, document extensions, ... www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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 = (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 www.ThesisScientist.com Each test involves: (i) Sampling (ii) CheckingIntroduction to Information Retrieval Sec. 19.5 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  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) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Statistical methods  Approach 1  Random queries  Random searches  Approach 2  Random IP addresses  Random walks www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Random URLs from random queries  Generate random query: how Not an English  Lexicon: 400,000+ words from a web crawl dictionary  Conjunctive Queries: w and w 1 2 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  This distribution induces a probability weight W(p) for each page. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Query Based Checking  Strong Query to check whether an engine B has a document D:  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  Frames  Redirects  Engine timeouts  Is 8word query good enough www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Advantages disadvantages  Statistically sound under the induced weight.  Biases induced by random query  Query Bias: Favors contentrich 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: Timeouts, failures, engine inconsistencies, index modification. www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Random searches  Choose random searches extracted from a local log Lawrence Giles 97+ or build “random searches” Notess  Use only queries with small result sets.  Count normalized URLs in result sets.  Use ratio statistics www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Queries from Lawrence and Giles study  adaptive access control  softmax activation function  neighborhood preservation  bose multidimensional system topographic theory  hamiltonian structures  gamma mlp  right linear grammar  dvi2pdf  pulse width modulation neural  john oliensis  unbalanced prior probabilities  rieke spikes exploring neural  ranked assignment method  video watermarking  internet explorer favourites  counterpropagation network importing  fat shattering dimension  karvel thornber  abelson amorphous computing  zili liu www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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 pages  Estimated use of metadata descriptors:  Meta tags (keywords, description) in 34 of home pages, Dublin www.ThesisScientist.com core metadata in 0.3Introduction to Information Retrieval Sec. 19.5 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.  E.g.: 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) www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 Random walks  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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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.  Nonuniform distribution  Subject to link spamming www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.5 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 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 DUPLICATE DETECTION www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 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., lastmodified date the only difference between two copies of a page www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Duplicate/NearDuplicate Detection  Duplication: Exact match can be detected with fingerprints  NearDuplication: Approximate match  Overview  Compute syntactic similarity with an editdistance measure  Use similarity threshold to detect nearduplicates  E.g., Similarity 80 = Documents are “near duplicates”  Not transitive though sometimes used transitively www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Computing Similarity  Features:  Segments of a document (natural or artificial breakpoints)  Shingles (Word NGrams)  a rose is a rose is a rose → aroseisa roseisarose isaroseis aroseisa  Similarity Measure between two docs (= sets of shingles)  Jaccard coefficient: SizeofIntersection / SizeofUnion www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 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 (sizeofintersection / sizeofunion) based on a short sketch Doc Shingle set A Sketch A A Jaccard Doc Shingle set B Sketch B B www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 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 i is as follows: D m  Let f map all shingles in the universe to 0..2 1 (e.g., f = fingerprinting) m  Let p be a random permutation on 0..2 1 i  Pick MIN p (f(s)) over all shingles s in D i www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Computing Sketchi for Doc1 Document 1 64 Start with 64bit f(shingles) 2 64 2 Permute on the number line with p i 64 2 64 2 Pick the min value www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Test if Doc1.Sketchi = Doc2.Sketchi Document 2 Document 1 64 64 2 2 64 64 2 2 64 64 2 2 A B 64 64 2 2 Are these equal Test for 200 random permutations:p , p ,… p 12 200 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 However… Document 2 Document 1 64 64 2 2 64 64 2 2 64 64 A B 2 2 64 64 2 2 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 Sizeofintersection / Sizeofunion www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Set Similarity of sets C , C i j C C i j Jaccard(C ,C ) i j C C i j  View sets as columns of a matrix A; one row for each element in the universe. a = 1 indicates presence of ij item i in set j C C 1 2  Example 0 1 1 0 1 1 Jaccard(C ,C ) = 2/5 = 0.4 1 2 0 0 1 1 www.ThesisScientist.com 0 1Introduction to Information Retrieval Sec. 19.6 Key Observation  For columns C , C , four types of rows i j C C i j A 1 1 B 1 0 C 0 1 D 0 0  Overload notation: A = of rows of type A  Claim A Jaccard(C ,C ) i j A B C www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 “Min” Hashing  Randomly permute rows  Hash h(C) = index of first row with 1 in column i C i  Surprising Property P h(C ) h(C ) Jaccard C ,C  i j i j  Why  Both are A/(A+B+C)   Look down columns C, C until first nonTypeD row i j  h(C) = h(C )  type A row i j www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 MinHash sketches  Pick P random row permutations  MinHash sketch Sketch = list of P indexes of first rows with 1 in column C D  Similarity of signatures  Let simsketch(C ),sketch(C ) = fraction of permutations i j where MinHash values agree  Observe Esim(sketch(C ),sketch(C )) = Jaccard(C ,C ) i j i j www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 Example Signatures S S S 1 2 3 Perm 1 = (12345) 1 2 1 Perm 2 = (54321) 4 5 4 C C C 1 2 3 Perm 3 = (34512) 3 5 4 R 1 0 1 1 R 0 1 1 2 R 1 0 0 3 R 1 0 1 4 Similarities R 0 1 0 5 12 13 23 ColCol 0.00 0.50 0.25 SigSig 0.00 0.67 0.00 www.ThesisScientist.comIntroduction to Information Retrieval Sec. 19.6 All signature pairs  Now we have an extremely efficient method for estimating a Jaccard coefficient for a single pair of documents. 2  But we still have to estimate N coefficients where N is the number of web pages.  Still slow  One solution: locality sensitive hashing (LSH)  Another solution: sorting (Henzinger 2006) www.ThesisScientist.comIntroduction to Information Retrieval More resources  IIR Chapter 19 www.ThesisScientist.com