personalized navigation for the web

7/28/2019 Personalized Navigation for the Web

1/3

ecomSys

Recom

men

derSystems

COMMUNICATIONS OF THE ACM March 1997/Vol. 40, No. 3 73

by the user. Over time, Siteseer learns each userspreferences and the categories through which theyview the world, and at the same time it learns for eachWeb page how different communities or affinity-based clusters of users regard it. Siteseer then deliverspersonalized recommendations of online content,

Web pages, organized according to each users folders.

Bookmarks (including Hotlists and Favorites) are adesirable mechanism for gathering preference infor-mation as they are already maintained by the user, andthus require no additional behavior for the purpose of

informing the recommendation system. In contrast toa click, which can be inadvertently done and rarelytakes much effort or investment, bookmarks are theresult of a very intentional act, something which(especially if the bookmark is placed in a folder) takessome degree of thought and effort, making them a lessnoisy input for inference.

Bookmarks also have specific limitations. A volun-tary survey of free response and multiple choice ques-tions, posted to various Usenet news groups anddrawing 40 respondents, indicated that users typi-cally bookmark fewer than half of the sites/pages they

find interesting, often because a site is easily accessi-

Personalized Navigation

for the Web

S

ITESEER IS A WEB-PAGE RECOMMENDATION SYSTEM THAT USES AN

individuals bookmarks and the organization of bookmarks within

folders for predicting and recommending relevant pages. Siteseerutilizes each users bookmarks as an implicit declaration of interest in the

underlying content, and the users grouping behavior (such as the place-

ment of subjects in folders) as an indication of semantic coherency or rel-

evant groupings between subjects.

In addition, Siteseer treats folders

as a personal classification systemwhich enables it to contextualize

recommendations in classes defined

Bookmarks can be a key component for gathering

preferential information.

James Rucker and Marcos J. Polanco

Siteseer:


2/3

ble through other means such asanother Web page or a search engine.In addition, users tend to bookmark

for wildly different reasons, rangingfrom genuine interest to a transientneed to return to a page. Finally,bookmarks either exist, or they dont.There is no partial bookmark thatwould indicate marginal interest, andthere is no way that bookmarks can beused to indicate a lack of preference,which an explicit feedback system canrequest.

Siteseer uses the findings of oneuser as implicit recommendations for

another based on the bookmarkeddiscoveries of a pool of reviewersqualified as trusted recommenders.The criteria for a reviewer being arecommender for another is straight-forward. Fundamentally, Siteseerlooks at each users folders and book-marks, and measures the degree ofoverlap (such as common URLs) of

each folder with other peoples folders,giving additional weight to URLsthat are more obscure, (that is, less

prevalent in user folders).1

The systemdoes not derive any semantic valuefrom the contents of the URLs, northe title of the folders; it uses the URLas a unique identifier and completelyignores the title. By using overlap ofcontents to determine folder similar-ity, Siteseer establishes concept simi-larity without relying on the titlesgiven to the folders.

Using this method, Siteseer formsdynamically defined virtual communi-

ties of interest, particular to each userand specific to each of the users cate-gories of interest. In our example, Johnhas a Vacation Spots folder which,like any other folder, is the basis for theformation of a virtual community. Inthis case, Marys Tropical Getawaysfolder has the highest overlap withJohns Vacation Spots, making herthe most qualified recommender.

http://www.dominique.orghttp://www.puertorico.org

http://www.cuba.org

http://www.bahamas.org

Islas Adjacentes

Tropical Getaways

http://www.vtourist.com

http://www.puertorico.org

http://www.pctravel.com

http://www.belize.org

http://www.canarias.org

http://www.costarica.org

http://www.vtourist.org

http://www.delorme.com

http://www.costarica.org

http://www.cuba.org


http://www.belize.org

Exotic Locations

Vacation Spots

http://www.cuba.org



http://www.alaska.org

http://www.canarias.org

http://www.france.org

http://www.bahamas.org


http://www.delorme.com

http://www.tahiti.orghttp://www.egypt.org


http://www.morocco.org

Caribbean Paradise

Joanne

John

Mary Christine

This figure shows four

qualified folders, selected from

a larger pool, which are partof the virtual neighborhood

surrounding Johns

Vacation Spots. Marys

Tropical Getaways is the

most qualified recommending

folder, if computed by simple

overlap. Also, note that


would be the highest

confidence recommendation,

given that it is coming from

Marys Tropical Getaways,and also because it is

contained in another

qualified recommenders

folder, Christines

Caribbean Paradise.

1This explanation is significantly simplified but con-

veys the main point; the mechanism for determin-

ing folder similarity is more complex and involvesadditional factors not mentioned.

74 March 1997/Vol. 40, No. 3 COMMUNICATIONSOF THE ACM


3/3

However, the converse is not true; for the virtual com-munity surrounding Marys Tropical Getaways,Johns Vacation Spots folder would not be the mostqualified. The most qualified recommender wouldeither be Christine, on the basis or her Caribbean Par-adise folder, or someone else, based on a more highly

correlated folder not shown in the diagram. By com-puting community membership relative to each folder(and not attempting to form a finite set of clusters),Siteseer imposes no category rigidity and is thus able toserve users even with relatively obscure or highly spe-cific interests.

Siteseer provides as recommendations those pageswhich have been bookmarked by the users virtualneighbors, giving preference to pages drawn fromfolders with the highest overlap as well as those heldwithin multiple folders in the neighborhood. Siteseercontextualizes its recommendations by delivering

them in the folder served as the basis for the discovery.Thus, in our example, recommendations coming fromusers in Johns Vacation Spots neighborhood will bedelivered within the context of his Vacation Spotsfolder.

While a novel recommendation and categoriza-tion system, Siteseer has intrinsic limitations becauseof its almost purely collaborative approach. Mostnotable is its inability to help the first-time user orone creating a new category for bookmarks, due tothe simple fact that until a community has been dis-

covered, there is no collective experience to leverage.Therefore, such users must first arrive at sites orpages by some other means, such as a search serviceor editorially built Web directory.

At press time, Siteseer had a fairly limited base ofjust over 1,000 users and performed best on foldershaving between 15 and 20 bookmarks, making rec-ommendations 88% of the time, with an average con-fidence factor of 18 out of 100 (which is partiallyindicative of folder overlap). For smaller and largersets, it performed fewer and lower-confidence recom-mendations. Overall, it was able to provide recom-

mendations for 60% of the folders.

James Rucker ([email protected]) is president andco-founder of San Francisco-based Imana, Inc., a pioneeringfirm in the field of collaborative filtering technology.Marcos J. Polanco ([email protected]) is chair andco-founder of Imana, Inc., San Francisco, Calif.

Permission to make digital/hard copy of part or all of this work for personal or class-

room use is granted without fee provided that copies are not made or distributed for

profit or commercial advantage, the copyright notice, the title of the publication andits date appear, and notice is given that copying is by permission of ACM, Inc. To copy

otherwise, to republish, to post on servers, or to redistribute to lists requires prior spe-

cific permission and/or a fee.

ACM 0002-0782/97/0300 $3.50

c

COMMUNICATIONSOF THE ACM March 1997/Vol. 40, No. 3 75

SIGs Announce Candidates

for Election

In accordance with ACM Bylaw 7, the followingSIGs will hold elections this spring: SIGAda,

SIGCAS, SIGCUE, SIGLINK.

ACM Policy and Procedures require that those

SIGs holding elections notify their membership of

candidates for elected offices. Below is a list of SIGs

that have submitted their slate of candidates.

In accordance with the SIG Bylaws, additional candidates may be

placed on the ballot by petition. All candidates must be voting

members of ACM and the SIG. Anyone interested in petitioning

must inform ACM Headquarters (Pat Ryan) and the SIG

Secretary of his/her intent to petition by March 20th.

SIGAda

ChairHarry F. JoinerBen Brosgol

Vice Chair forMeetings andConferences

Currie ColketDavid Harrison

Vice Chair forLiaison

S. Ron Oliver

Steve Grimaldi

TreasurerBard CrawfordChad Bremmon

SecretaryTerry WestleyJohn W.

McCormick

InternationalRepresentativeJean-Pierre RosenKarlotto Mangold

SIGCAS

ChairC. Dianne MartinTom Jewett

Vice ChairDon GotterbarnofKeith Miller

Secretary/TreasurerSuzanne WeisbandHerman Tavani

SIGCUE

ChairJohn T. GorgoneJames K. Hightower

Vice ChairRuth Anne RossMarie-Michele

BouletSecretary/Treasurer

John LawsonTerris B. Wolff

SIGLINK

ChairMarc NanardMark Bernstein

Vice Chair (UnitedStates)

H. Van DykeElli Mylonas

Vice Chair (Europe)Franca GarzottoKaj Gronbak

SecretaryFrank ShipmanSteven J. DeRose

TreasurerJim RosenbergUffe Wiil

personalized navigation for the web

Documents