Using Link Analysis to Identify Aspects in Faceted Web Search

Using Link Analysis to Identify Aspects in Faceted Web SearchChristian Kohlschütter L3S/University of Hannover Deutscher Pavillon,Expo Plaza1 30539Hannover,Germany kohlschuetter@l3s.dePaul-Alexandru ChiritaL3S/University of HannoverDeutscher Pavillon,Expo Plaza130539Hannover,Germanychirita@l3s.deWolfgang NejdlL3S/University of HannoverDeutscher Pavillon,Expo Plaza130539Hannover,Germanynejdl@l3s.deABSTRACTAre you looking for information about bush in the context of Politics or Gardening,with focus on government sources or news agencies?Faceted Search can improve your web search by o ering drill-down options towards such speci c aspects,particularly in controlled document collections with rich human-edited annotations.The Web,for the most part, however lacks appropriate metadata;only a small fraction is covered by taxonomies or folksonomies.In this paper,we show that the Web's intrinsic link structure is su cient to enable an automatic faceting on such a large and dynamic document collection,not being limited to topic detection only.We present an e cient method for identifying facets within full-text search.For each page,we allot fuzzy mem-bership values based on Personalized PageRank to all facets, then we compare them in order to derive context informa-tion.A prototype implementation is available online.1.INTRODUCTIONThe basic principle of web search is to nd a set of pages ranked by relevance to an idea or concept the user has in mind.These concepts are usually speci ed by keywords that must occur in any document deemed relevant.Obviously,it is often di cult to nd terms which precisely separate rele-vant from irrelevant pages,because they are either ambigu-ous or can at least be seen from di erent perspectives.While modern search engines may satisfy the average user's needs by focusing on general importance(i.e.,a keyword search for bush primarily returns pages about George W.Bush), as soon as other aspects are concerned,the only way out is to specify additional keywords(i.e.,do a search for bush gardening to focus on horticultural issues)and hope that these auxiliary terms are not too restrictive.But the more aspects are speci ed as terms(such as scienti c article and British ),chances are high that many highly relevant docu-ments are ltered out because not all terms are matched[10]. The emerging paradigm of Faceted Search[6,7,10,19]may help in this situation.Here,the system automatically de-Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on thefirst page.To copy otherwise,to republish,to post on servers or to redistribute to lists,requires prior specific permission and/or a fee.SIGIR’2006Workshop on Faceted Search,Aug10,2006,Seattle,W A,USA. Copyright2006ACM X-XXXXX-XX-X/XX/XX...$5.00.termines possible aspects from the result set and presents them to the user,modeled as categories of orthogonal di-mensions like topic,cultural background or target audience and nally enables the user to iteratively narrow( drill-down )the search until he is satis ed with the results.These facets 1are represented as metadata,they do not interfere with full-text keywords as opposed to text-based clustering approaches[10].While faceted search works well in situations where facet metadata is annotated to the documents like in library sys-tems or enterprise search applications(e.g.,restaurant search), on dynamic,large scale collections of heterogeneous docu-ments such as the Web,the faceted classi cation of pages and the identi cation of facets within search results still are unresolved problems.The lack of annotations and the om-nipresence of noise within this collection hamper a direct adoption of enterprise faceted search.In this paper,we focus on the question how we can still utilize the few annotations from taxonomies or folksonomies like ODP or for faceted web search.We present an e cient method for automatically identifying facets in web search results solely by link analysis.First experiments have shown that we get a high precision of the suggested faceted classi cations without requiring additional data structures or categorization algorithms.The paper is organized as follows.In Section2,we give a short overview of faceted search,as well as of taxon-omy/folksonomy based web page tagging systems and the Personalized PageRank algorithm.In Section3,we describe an approach to classify web pages through PageRank-based link analysis.Based thereupon,Section4covers our algo-rithm for improving full-text search results by identifying facets with respect to the given keywords.In Section5,we describe our demo application and show the results of a pre-liminary evaluation.Section6concludes with a discussion of on-going and future work.2.RELEV ANT BACKGROUNDFaceted search.Faceted search is a relatively young research area,and thus there exist only few approaches to tackle the problems it raises,especially when generalizing this kind of search for the entire web environment.Facets allow for di erent views on the result set,which can be obtained through a specialized user interface[19],thus en-abling the user to choose di erent possible starting paths 1Note:In this paper,we use the term facet as a synonym to aspect or category ,whereas we call the orthogonal axes facet dimensions .for the exploration of the collection.The Flamenco Sys-tem2for example allows for both searching and browsing an image collection from various perspectives,such as gender, country of a liation for Nobel prize winners.There exist at least two types of facets:(1)Hierarchical and(2)Flat. The concept of Hierarchical Faceted Categories(HFC)was introduced by Hearst et al.in[11,19],and it relies on a set of category hierarchies(one per facet),built manually in advance.Thus,they classify the documents available in a collection space according to each of the hierarchies sep-arately,producing several navigation paths across the same set of points.Inherently,users will nd their sought docu-ments faster,as more routes towards them exist.The main approach we discuss in this paper,i.e.,using page topic(s) as one facet for categorizing web results,also falls in the HFC type.The second type of facets is a at one,in which there is no clear relation between the elements generated within the same facet dimension.We are not aware of any prior work creating this kind of facets from within textual web documents.Large scale taxonomies/folksonomies.One of the largest e orts to manually annotate web pages is the Open Directory Project(ODP)3.Over70,000editors helped to categorize more than5million web sites into almost600,000 hierarchical categories describing web sites'topics;however, these pages still make far less than0.1percent of the pub-licly accessible web.While in ODP,the taxonomy is clearly split into16root categories,such as Business,Computers or Sports, folksonomical organized platforms like 4 allow users to annotate arbitrary tags to web pages rather than using a xed taxonomy[8,14].Since both approaches allow to classify pages to more than just one facet(category or tag),faceted search could easily be implemented for the pages annotated in these collections,but not for the whole web graph.Personalized PageRank.In this paper,we focus on one speci c approach to exploit directly annotated information for supporting faceted search on large-scale web graphs:the so-called Personalized PageRank[15],a special version of the well-known PageRank web page popularity measure[15]. It promises an improvement in ranking with respect to in-dividual interests/facets,which are speci ed through a set of relevant pages.Regular PageRank computes a general importance score r(p)for every page p of the web graph G=(P,L),forming the PageRank vector r.It models the random surfer,who visits the web graph in a rather un-biased way:it starts from an arbitrarily chosen page and follows one random link from one page to another( walk ) until it gets bored,then it jumps to another random page in the graph,and so on.In Personalized PageRank,the jump now only addresses a given set of pages to which the compu-tation should be biased.A common representation of both approaches(random/biased)is the following:∀t∈P:r B(t)=(1−d)·πB(t)+d·X(s,t)∈L r B(s)|L(s)|(1)Here,for a given biasing set B⊆P the PageRank score 2/3/4/r B(t)for a page t is de ned recursively,based on a weighted5 linear combination of jump and walk action.The second summand describes the likelihood that the surfer reaches the page t by walking,which commonly is expressed by the sum of the ranks of all pages s linking to t,each divided by their number of outgoing links|L(s)|.Finally,πB(t)describes the likelihood that the surfer reaches page t from a jump.In the case of regular( unbiased )PageRank,we treat all pages of the web graph as being contained in B,so:πP(t)=|P|−1, in general it is:πB(t)=(1|B|i t∈B0otherwise(2)The distribution of scores in the rank vector r B follows the power law and thus is scale-free.Therefore multiplying a PageRank score with the keyword-based document rele-vance score W(d,q)(e.g.,based on the T F×IDF measure) causes important pages(with respect to the given biasing set B)to gain higher scoresρB.Note that the number of retrieved documents does not change when switching to an-other biasing set,since no facet-speci c keywords are added. Instead the document order is changed:ρB(d,q)=W(d,q)·r B(d)(3) Personalizing PageRank is in fact just a biasing towards an individualized set of scores.As a non-user speci c ap-proach,Haveliwala's Topic-Sensitive PageRank[9]makes use of linear combinations of just16Personalized PageR-ank vectors,each biased on one of the16top categories of ODP respectively.The weights for the linear combination are derived from each topic's term vector(hence,access to full-text information is required):r B =Xβ[ωβ rβ](4)Other,more recent approaches to generate personalized rankings do already exist.Qiu and Cho[16]for example also extend Topic-Sensitive PageRank to contain user spe-ci c weights for combining the16biased vectors.They learn these weights by applying machine learning on user's click history.Aktas et al.[1]have successfully applied Personal-ized PageRank with respect to domain name features(coun-try and generic TLD topic assignment).Finally,Jeh and Widom[12]and more recently Sarlos et al.[17]have inves-tigated the means to compute Personalized PageRank in a scalable way for a large set of users.They both exploited the idea of decomposing the biasing set into small sets with a single non-zero entry followed by a linear combination of the resulted Personalized PageRank vectors.3.WEB PAGE CLASSIFICATIONUSING PERSONALIZED PAGERANK Several aspects of a single page may be considered rele-vant for faceted search;they can be ordered into facet di-mensions like thematical coverage,language,age etc.To automatically retrieve such facets,classi cation algorithms can be applied to receive the set of relevant facets M p for each page p in the web graph.If possible,a relevance mea-sure is assigned to each facet,resulting in a relevance vector 5d usually is around0.85in literature.m p whereas each dimension represents one of all available facets.In classical IR,to compute m p text-based mod-els are used,such as Naive Bayes or LSA.For the Web, one can consider the hyperlink structure of the web as an-other source of classi cation power.Several approaches ex-ist which either combine link structure and textual informa-tion or even attempt to derive a classi cation from the web graph only[3 5,13,18].While they may perform very well, in all cases the only purpose of the proposed algorithms is to generate a classi cation being not related to PageRank; the imposed additional payload(computational and storage requirements)to the search engine system may not be un-derestimated.In this section,we present an e cient method which re-uses the rank vector of Personalized PageRank for classi cation,avoiding such overhead .We are convinced that Personalized PageRank vectors themselves provide su cient information for classifying pages towards arbitrary facets.Because of the skewed jump in the PageRank computation,pages within the corresponding bi-asing set tend to get higher scores than pages outside.Via the link structure also pages being outside but nearby the biasing set(in terms of link hops)gain high scores[15].As-suming that pages mostly link to somehow related pages, there consequently must be a direct relationship between a page's facet-speci c PageRank score r f(p)and its member-ship m p(f)to that facet.m p(f)∝r f(p)(5) We can treat the values as quali able membership to the facet f(member of the set of available facets F);values above the average scoreµ≈1.0are treated as positive membership,values below as negative membership(i.e., non-relevant to the facet).One may assume that the mem-bership value is equal to the rank value,or that a high rank score automatically means a high relevance to the facet(as suggested in[9]).However,since there also is a fair amount of pages of general importance/interest which may also have a high score[4],directly proportional relationship is too imprecise.We therefore introduce the facet membership uncertaintyαf(p)for the page p:the more facets are pos-sible the less reliable the membership statements are.Since facet dimensions are orthogonal,this uncertainty depends on the number of facets in the same dimension dim as the considered facet f.In case that no facets could be deter-mined,the uncertainty is in nite:G f(p)={g∈F|dim(g)=dim(f)∧r g(p)≥µ}(6)αf(p)=(∞i G f(p)=∅|G f(p)|otherwise(7)To account for the scale-free nature of the rank score,αf should in uence r f exponentially:m p(f)=log r f(p)f(8)m p(f)now also quantitatively re ects positive or negative membership certainty6.Having said that,a value around 6We shall avoid the term probability;neither r f(p)norαf(p) are bounded.zero does not implicate that a page does not relate to a spe-ci c facet it is just unclear.This case applies to commonly linked pages like the Acrobat Reader download page,the ph-pBB bulletin board system website and others.On the other hand,specialty pages clearly bene t from this classi cation.4.IDENTIFYING FACETS FROMWEB SEARCH RESULTSWhen submitting a text query q on a PageRank-supported web search engine,the top-k of the returned result list rep-resents the k most relevant results for the query terms with respect to general importance(for Unbiased PageRank)or to speci c facets(for Personalized PageRank).For a su -ciently large k,this list could be regarded as a sample set S q of the whole result set D q;another approach to deter-mining S q,taking care of strongly over-represented facets, is shown in[2].Now,we can compute the facet membership vector m p for all pages p∈S q.From this,we can derive the facet membership vector m S,q,which represents the facet memberships for the given top-k subset S q of results for the query q.In other words,this vector represents the facets which are deemed to be contained in the full result set D q. In order to compute m S,q we use the following procedure: For each facet membership value(in each vector m p,for all pages p∈S q)which exceeds a certain thresholdϑ,we increment the corresponding value m S,q(f)by one; nally,m S,q is normalized to values between0and1.0:m S,q(f)=|{p∈S q|m p(f)≥ϑ}||S q|(9)As an improvement to facet diversity,we suggest to com-pute the membership values not only from one sample set but from multiple PageRank result sets.For example,with the ODP taxonomy,we start with sampling the top-k results from Unbiased PageRank as well as all Personalized PageR-ank vectors based on the taxonomy's16top categories.If the user has already chosen an ODP-topic as a facet,we take the corresponding child categories instead.Another way to extend the set of facets,which can also be applied to non-hierarchical facets,is to start from a custom seed of facets and then iteratively determine relevant facets using the formula above and extend this seed by newly de-tected facets D(see algorithm1).The user can now select from the facets given in D for the re nement of his search.If more than one facet is selected,Algorithm1Iterative identi cation of facetsλ=Membership threshold for newly discovered facets D←∅D ←Facet seedi max←Maximum number of iterationsi←0while i<i max∧D \D=∅<doi←i+1D←D∪DD ←{g∈F|m q,D(g)≥λ};m q,D(f)=Pd∈D|{p∈S d,q|m p(f)≥ϑ}|Pd∈D|S d,q|end whileFigure1:Screen-shot of the prototypea Personalized PageRank is used based on the linear combi-nation of the facets'rank vectors,as described in Section2. As opposed to Topic-Sensitive PageRank,we think it is not necessary to derive the combination weights from textual document statistics.Instead,we may propose pre-de ned weights on the level of facet dimensions(e.g.,prefer topic over origin),which can be altered by the user;facets of the same dimension may safely be equally balanced.5.IMPLEMENTATION AND EV ALUATION Right now,we have implemented facet detection for only one dimension,the topical membership of page,based on the ODP taxonomy.A screen-shot of the prototype is de-picted in gure1;the demo can be tested online,at http: //www.l3s.de/facetedsearch.We are currently extending the setup by two more dimensions,domain name features and type of information source(academia,business,govern-ment,media,private etc.).Our search engine prototype is based upon the Lucene7 information retrieval software library,which we extended to hold multiple PageRank scores per document.We used a collection of9.2million documents which we have retrieved in2005using the Heritrix crawler8.The crawler's strategy was to nd pages from a variety of topics.We performed a broad crawl based on a seed of the pages from four of the16top-level categories of ODP,Business,Computers, Sports and Recreation.From them about one hundred bi-asing sub-categories were randomly chosen.This selection process was executed as follows:For each of the four top categories,three subcategories were randomly picked;then, for one of them,we again randomly took three subcategories and so on,until no deeper levels were available;almost all paths ended at level6(with level1being one of the ODP root categories).We then computed the corresponding100+ Personalized PageRank vectors(plus the Unbiased PageR-ank)using the pages residing in each of these categories (including sub-categories)as biasing sets.7/8/Any keyword search is conveyed using Unbiased PageR-ank by default;however the user can manually choose any available biasing set.The search results are classi ed on-line using equation8;since the classi cation is fuzzy and the op-eration can be performed quickly(no access to textual data is required),we decided not to store this information stat-ically in the index.For the top-k results(k=100per de-fault),we identify the contained facets having a membership certainty of at leastϑ(0.1per default).Both parameters can be adjusted by the expert user.The detected facets are presented next to the document descriptions;facets based on ODP topics are shown both as a tree as well as a sorted list,along with the facet membership value from algorithm 1.The user can choose from these facets by clicking on the corresponding facet name(facets with a membership of at least25%are highlighted).The search is then repeated using the corresponding Personalized PageRank. Unbiased PageRank is treated as a special facet.It may also receive a membership value.Intuitively,this facet can be regarded as general importance.If no other facet has been retrieved for the keyword query,we conclude that we lack the proper Personalized PageRank.From this perspec-tive,the membership to general importance can be inter-preted as uncertainty as well.We performed a preliminary evaluation of the facet de-tection algorithm.We determined a set of phrase queries which are relevant to speci c ODP categories(33in our test set,5keywords each)and searched for these phrases with our system(i max was1).In order to avoid falsi ed results,we did not distill these queries from our own data set,but utilized Google's AdWords service9(we randomly selected5suggested queries;an excerpt is shown in the ap-pendix).Then we compared the category recommendations from AdWords with the topical facet recommendations from our system.Exact phrase search returned results in155out of the165cases.Only in13cases,the algorithm did neither retrieve the desired category nor any ancestor category(this makes a precision of91.6%).9https:///6.CONCLUSIONS&ONGOING WORKIn this paper we proposed to drill-down from hierarchical faceted search and to simply re-order web search results ac-cording to a speci c category within a facet using Personal-ized PageRank.Our current implementation covers output categories for the topic facet,thus enabling users to on-the- y switch to result rankings according to PageRank biased on some Open Directory topic.We have also proposed a new simple technique to infer the most relevant topics as-sociated to a user query.Experimental results have shown this approach to yield precise identi cation of facets.We are currently implementing the generation of a sec-ond type of categories,according to the source type facet. This would enable biasing PageRank on news pages,on user home pages,on university sites etc.For completeness,we also think about adding facets that can be directly retrieved from the documents themselves,such as document size and language.We plan to nish this before giving the demo at the workshop.7.REFERENCES[1]Mehmet S.Aktas,Mehmet A.Nacar,and FilippoMenczer.Personalizing pagerank based on domainpro les.In WEBKDD'04:Proceedings of the sixthWEBKDD workshop:Webmining and Web UsageAnalysis(WEBKDD'04),Seattle,USA,pages83 90,August2004.[2]Aris Anagnostopoulos,Andrei Z.Broder,and DavidCarmel.Sampling search-engine results.In WWW'05: Proceedings of the14th international conference onWorld Wide Web,pages245 256,New York,NY,USA,2005.ACM Press.[3]Pavel Calado,Marco Cristo,Marcos Andre;Goncalves,Edleno S.de Moura,BerthierRibeiro-Neto,and Nivio Ziviani.Link-based similarity measures for the classi cation of web documents.J.Am.Soc.Inf.Sci.Technol.,57(2):208 221,2006. [4]Soumen Chakrabarti,Byron Dom,and Piotr Indyk.Enhanced hypertext categorization using hyperlinks.In SIGMOD'98:Proceedings of the1998ACMSIGMOD international conference on Management of data,pages307 318,New York,NY,US,1998.ACMPress.[5]Je rey Dean and Monika R.Henzinger.Findingrelated pages in the World Wide puterNetworks(Amsterdam,Netherlands:1999),31(11 16):1467 1479,1999.[6]William Denton.How to make a faceted classi cationand put it on the web./ library/facet-web-howto.pdf,November2003. [7]William Denton.Putting facets on the web:Anannotated bibliography.http:///library/facet-biblio.html, October2003.[8]Scott Golder and Bernardo A.Huberman.Thestructure of collaborative tagging systems.Technicalreport,Information Dynamics Lab,HP Labs,2005. [9]Taher H.Haveliwala.Topic-sensitive PageRank.InProc.of the eleventh International Conference onWorld Wide Web,pages517 526.ACM Press,2002.[10]Marti A.Hearst.Clustering versus faceted categoriesfor information mun.ACM,49(4):59 61,2006.[11]Marti A.Hearst,Ame Elliott,Jennifer English,Rashmi R.Sinha,Kirsten Swearingen,and Ka-PingYee.Finding the ow in web site mun.ofthe ACM,45(9):42 49,2002.[12]Glen Jeh and Jennifer Widom.Scaling personalizedweb search.In WWW'03:Proceedings of the12thinternational conference on World Wide Web,pages271 279,New York,NY,USA,2003.ACM Press. [13]Qing Lu and Lise Getoor.Link-based textclassi cation.Text-Mining&Link-Analysis Workshop TextLink2003,2003.[14]Adam Mathes.Folksonomies-cooperativeclassi cation and communication through sharedputer Mediated Communication-LIS590CMC,Graduate School of Library andInformation Science,University of IllinoisUrbana-Champagin,December2004.[15]Lawrence Page,Sergey Brin,Rajeev Motwani,andTerry Winograd.The PageRank citation ranking:Bringing order to the web.Technical report,Stanford Digital Library Technologies Project,1998.[16]Feng Qiu and Junghoo Cho.Automatic identi cationof user interest for personalized search.In Proc.of the 15th international World Wide Web conference,2006.[17]Tam�s Sarl�s,Andr�s A.Bencz�r,K�rolyCsalog�ny,D�niel Fogaras,and Bal�zs R�cz.To randomize or not to randomize:Space optimalsummaries for hyperlink analysis.In Proc.of the15th international World Wide Web conference,2006. [18]Yiming Yang,Sean Slattery,and Rayid Ghani.Astudy of approaches to hypertext categorization.Journal of Intelligent Information Systems,18(2-3):219 241,2002.[19]Ka-Ping Yee,Kirsten Swearingen,Kevin Li,andMarti A.Hearst.Faceted metadata for image searchand browsing.In Proc.of the SIGCHI Conf.onHuman Factors in Computing Systems,pages401 408,2003.Table1:Sample keywords from Google AdWordsBusiness:information technology;market share;strategic planning;supply chain management;swot analysis Business/Textiles and Nonwovens/Textiles/ Carpets:rugs; oor covering;persian carpet;oriental weavers;hardwood oorComputers:laptops;workstations; at screen;second hand computers;midwest microComputers/Internet/Searching:search the web;web browser;front crawl;search engines;internet searching Recreation:recreation jobs;parks and recreation;nude recreation;recreation and leisure;lake mead recreation Recreation/Travel/Travelogues:travelogue game; travel diary;travelogues;travel adventures;travel phots Sports:n ;athletics;stadiums;formula one;basketball Sports/Soccer/UEF A/England/Women:women soccer uk;football girls;ladies soccer;girls fc;manchester ladies。