Structure-aware web transcoding for mobile …web.eecs.umich.edu/~zmao/eecs589/papers/transcode.pdfStructure-Aware Web Transcoding for Mobile Devices Transcoding heuristics that consider

Yonghyun Hwang,Jihong Kim, and Eunkyong SeoSeoul National University

Structure-AwareWeb Transcoding for Mobile Devices

Transcoding heuristics that consider a Web page’s structure

and the relative importance of Web components make

complex Web pages accessible from handheld devices.

As both the number of mobile usersand users’ reliance on the Webgrows, so does the need for Web

access from handheld devices.1 The cur-rent disparity between such devices’available computing resources and theresources required for smooth Webbrowsing makes it difficult and unpleas-ant to access Web pages with them. Tonavigate complex Web pages with ahandheld device, a user must scroll downand across the page many times. Further-more, although handhelds are becomingmore powerful, few can satisfactorilyhandle multimedia data.

Two general solutions to this disparityexist: manual and automatic reauthoring.In manual reauthoring, Web authors pre-pare multiple versions of a Web page tar-geted to resource profiles of various plat-forms, including the Wireless ApplicationPlatform (www.wapforum.org).2,3 Althoughthis approach can produce high-qualitypages for specific devices, it assumes a Webauthor will both be available to reauthor

the pages and will know what pages userswill want to access. Because no one canpredict how Web surfing will progress (dur-ing a mobile Web search session, for exam-ple), this approach severely limits the num-ber of Web pages accessible via handhelddevices.

In automatic reauthoring,4,5 a trans-coding module transparently convertsindividual pages, making all Web pagesaccessible. Although attractive in theory,this approach is not yet widely usedbecause the transcoding quality is poor.Systems such as Pixo (www.pixo.com/products/products002.htm) generate al-most unusable pages from complex Webpages with large nested table structures.Existing automatic reauthoring tech-niques ignore the relative importance ofWeb page components — basic units fortranscoding — which can be extractedthrough syntactic analysis of the page’sHTML source code. Furthermore, becausemost existing transcoding techniques arelocal transforms, they do not consider the

14 SEPTEMBER • OCTOBER 2003 Published by the IEEE Computer Society 1089-7801/03/$17.00©2003 IEEE IEEE INTERNET COMPUTING

Mob

ile

Web

Acc

ess

Web page’s overall structure when making deci-sions. (For a description of current techniques, seethe sidebar, “Existing Transcoding Heuristics.”)Recent work by Jinlin and colleagues6 considers aWeb page’s overall layout during the transcodingprocess, but does not reflect the relative impor-tance of Web components.

Our long-term goal is to develop a high-qualitysyntax-based Web transcoding system that allowsuniversal access to Web pages without manualreauthoring. We focus on structure-aware trans-coding heuristics, which preserve the original Webpage’s underlying layout as much as possible. Theproposed heuristics extract the relative importanceof Web components from an intelligent syntacticanalysis. Like other transcoding approaches, oursdoes not support executable content such as Java,Javascript, and Flash. To evaluate our heuristics’effectiveness, we performed subjective quality eval-uations in Web-based experiments.

Transcoding Framework We first describe a transcoding framework, which isuseful in understanding the strengths and weak-nesses of transcoding heuristics in a unified setting.

Web Page Representation We represent each Web page using a modified treestructure. This lets us efficiently analyze pagesbecause it represents various page layouts (forexample, nested table structures) and expresses theinterrelationship among components. Figure 1ashows a tree-based internal representation usinga sample Web page.

The tree-based data structure has two nodetypes and two edge types. Context nodes (whiteboxes in Figure 1a) contain attributes of the cor-responding HTML structure, such as the HTML tag,the estimated screen real estate the HTML struc-ture requires, and other heuristic-specific data.Terminal nodes (shaded boxes in Figure 1a) includeWeb content to be displayed, such as text blocks

and images. The tree’s solid edges point to the(possibly nested) substructures, and the dashededges represent the sibling relationships betweenconnected nodes.

Problem Formulation A Web transcoding technique H generally consistsof two functions: a grouping function and a sum-marizing function.

The grouping function Hg divides a Web pageinto several subgroups, forming Web components.Given a tree representation Tw of a Web page W,Hg partitions Tw into a set of subtrees.

The summarizing function Hs includes threesubtasks:

• deciding which subgroups will be reduced tohyperlinks in the transcoded pages;

• choosing representative phrases for the elidedsubgroups; and

• modifying the tree-based representation toreflect decisions made in the first two steps.

Consider, for example, the outlining transform O,4

which separates section headers from subsequenttext, as Figure 1 shows. The grouping function Og

identifies sections based on section headers in theWeb page (Figure 1a). The summarizing function Os

leaves the section headers as hyperlinks, hiding thetext blocks behind them (Figure 1b).

Given a tree representation Tw of a Web page Wand a set of n transcoding heuristics {H1, ..., Hn}, theWeb transcoding problem aims to find a sequence S = Hin º Hin–1

º ... º Hi3º Hi of transcoding heuristics

such that the transformed pages most efficiently sat-isfy a given quality metric. (For existing transcod-ing heuristics, see the sidebar on p. 18.)

Structure-AwareTranscoding Heuristics To improve the presentation of transcoded Webpages on a mobile device, transcoding heuristics

IEEE INTERNET COMPUTING http://computer.org/internet/ SEPTEMBER • OCTOBER 2003 15

Web Transcoding

Figure 1. Tree-based representation of a Web page. (a) The outlining transform’s grouping function divides Web componentsinto subtrees; (b) the summarizing function chooses subtrees to be elided and modifies the tree to reflect these decisions.

(a)

Text 1

HTML

BODYTITLE

H1 P

Section 1 Text 1

H1 P

Section 2 Text 2 Text 2

HTML

BODY

H1 P H1 P

HTML

BODY

Section 1

A

HTML

BODY

A

(b)

TITLE

DocumentDocument

Section 1 Section 2Section 2

should preserve the layout of the original Webpages as much as possible. We introduce two newheuristics, the generalized outlining transform andthe selective elision transform, which preserve thestructure of Web pages during transcoding. Bothexploit common layout characteristics of complexWeb pages. They use more general grouping func-tions than other transcoding heuristics while con-sidering the page’s overall layout. (For a moredetailed description, see our earlier work.7)

Generalized Outlining Transform Complex Web pages generally use multiple repeat-ed layout patterns. This is a characteristic thatother art mediums share. For example, mostgraphic artists know that large posters should notuse more than four colors or patterns; to deliverthe message more effectively they should use a fewpatterns repeatedly.

Figure 2a shows three repeated layout patterns ina sample Web page. If we can identify these patterns,we can group them like the outlining transformgroups header and text components. To detectrepeated layout patterns, we need a general groupingcondition. For example, we cannot use the outliningtransform to group the repeated patterns in the sam-ple page because the patterns do not use section

headers. The generalized outlining transform, how-ever, can identify these patterns using a generalgrouping function that transcodes the Web page to asequence of smaller pages, as Figure 2b shows.

Grouping function. The generalized outliningtransform’s grouping function depends on repeat-ed layout patterns, rather than specific tag combi-nations. Although repeated layout patterns areeasy to identify visually, they are difficult to detectin a tree-based representation. Successive Webcomponents often have the same layout but dif-ferent tree representations because they containsignificantly different content. Consider, for exam-ple, the five leftmost table cells in Figure 2a. AsFigure 2b shows, the cells form a repeated layoutpattern although the bottom cell in Figure 2a con-tains an image while the top four cells containhyperlinks only.

To find all the repeated layout patterns in a Webpage W and assign them to groups, the groupingfunction traverses W’s tree-based representationTW in postorder. Whenever it meets a context noderepresenting a structural tag (such as <ul> or<table>), it executes the prefix pattern-matchingalgorithm to find a repeated layout pattern from asubtree rooted with the current context node, orstructural context node.

To find a repeated layout pattern for each sub-tree T′ rooted with a structural context node, wesolve the prefix pattern-matching problem on thestring representation of subtree T′. Given the tree-based representation T′, we convert T′ to a stringrepresentation ST′, consisting of symbols represent-ing various HTML tags and text blocks in the Webpage. For example, we would convert the sourcecode <a>WebAlchemist</a><br> is a struc-ture-aware transcoding system to the string,TA Text Etag TBR Text. TA corresponds to <a>, TBRrepresents <br>, Etag represents all closing tags suchas </a>, and Text indicates a text block.

We search for frequently occurring substringsof ST′, which are candidates for a repeated layoutpattern. Because we match prefixes only, the dif-ferent content types do not affect matching results.The suffixes in a string representation denote theWeb component’s content type.

To find a repeated layout pattern, we constructa string tree for a given string representation. Eachstring tree edge represents a string element, whichsymbolizes an HTML tag or text block. Each noderecords the number of times the correspondingsubstring occurs in the given string representation.We construct the corresponding substring by fol-

16 SEPTEMBER • OCTOBER 2003 http://computer.org/internet/ IEEE INTERNET COMPUTING

Mobile Web Access

(a)

(b)

Repeatedlayout

pattern #1

Repeatedlayout

pattern #2

Repeatedlayout

pattern #3

Figure 2. An example of the generalized outlining transform. (a) Thetransform identifies three repeated layout patterns in a Web pageand (b) transcodes the page into a sequence of smaller pages.

lowing the edges from the root to the node.Figure 3 illustrates how we construct a string

tree, using repeated layout pattern #2 in Figure 2as an example. When we insert the string TA TextEtag Text TBR into a string tree, the string followsan edge TA, if it exists, under the root. If not, wecreate a new node and set its node counter to 1.This node is linked with the root, and we associ-ate the edge between the root and the new nodewith TA. We similarly insert Text, followed by Etag,Text, and TBR. As we insert each character of thisstring, the corresponding node counter incrementsby 1. The string tree in Figure 3 shows that thestring TA Text Etag Text TBR repeats six times.

After building the string tree, we can easily findrepeated substrings through tree traversal. We col-lect all substrings appearing more frequently thana given threshold as candidate repeat substrings.We then select the final repeated substring usingtwo rules:

• Rule 1. Select substrings starting with a struc-tural tag or a hyperlink.

• Rule 2. Select the longest of the selected sub-strings.

Finding a repeated substring takes O(n2) opera-tions, where n is the number of nodes in a subtreerooted by a given node.

We compute the repeated layout pattern fromthe selected repeated substring. Because we onlymatch the repeated substring’s prefix, the selectedsubstring defines the beginning portion of arepeated layout in a Web component. The group-ing function combines all the components locat-ed between the beginning of the current patternand the beginning of the next.

Summarizing function. After the grouping func-tion identifies the repeated layout patterns, thesummarizing function decides which repeated lay-out patterns to elide. It converts the matched pre-fixes of the selected layout patterns into hyper-links, hiding the rest of the content behind them.For example, it replaces repeated layout pattern #1in Figure 2a with five hyperlinks using thematched prefixes (category names such as “LDPworldwide”) in Figure 2b.

Because complex Web pages typically have sev-eral repeated layout patterns, the summarizing func-tion must select the appropriate patterns fromamong many candidates. We introduce two newparameters, shrinking factor and information densi-ty, to assist the selection process. The shrinking fac-

tor is the ratio of a Web component’s display sizeafter it is transcoded to its size before it is transcod-ed. The summarizing function generally favorsrepeated layout patterns with high shrinking factors.

Information density refers to the amount ofcontent accumulated in a Web component, indi-rectly representing the component’s importance.The summarizing function does not select repeat-ed layout patterns with high information densitybecause eliding such semantically important por-tions will degrade the transcoding quality.

Each Web component starts with an informationdensity of 1. In the transform’s current version, aWeb component’s importance is determined bysyntactic attributes such as font size, table cellwidth, cascading style sheets (CSS), and the numberof already elided subcomponents. The informationdensity increases by 1 if the font size or the tablecell width is larger than a predefined thresholdvalue. (This is based on the assumption that HTMLauthors commonly use larger fonts or wider tablecells to emphasize more important components.) Ifa Web component includes previously elided sub-components, its information density also increas-es by 1, which prevents the same component from


Web Transcoding

Figure 3. A simple example of string tree construction. Each edgerepresents a string element; inside each node is the number of timesthe substring is repeated.

TextTA

Text

Etag

Text

Text

Etag

TA

TA

TBR

TBR

TBR

.....

.....

.....

.....

Root

6

6

6

6

6

6

6

6

5

5

5

5

being elided repeatedly, which makes it difficult topreserve the Web page’s original layout. The Webcomponents in the lower right corner of Figure 2ahave an information density of 3 because they useboldface text in a wide table cell.

After assigning a shrinking factor and an infor-mation density to each Web component, the sum-marizing function classifies the components intofive categories based on these numbers.7 For exam-ple, if a component has a high shrinking factor butlow information density, the summarizing functionelides it. On the other hand, if the component hasa high information density, the summarizing func-tion does not elide it because it is important.

Selective Elision Transform Many popular Web sites use complex tables toorganize their content structurally. Properlytranscoding such complex Web pages for handheld

devices is a challenge. Existing transcoding heuris-tics such as the table transform (see the sidebar)often destroy the original table structures. Once thetable structure is broken, it is difficult to under-stand the author’s intent.

The selective elision transform attempts to pre-serve table structure as much as possible: in addi-tion to table attributes such as cell width, thistransform uses syntactic attributes such as fontsize to decide whether to elide a table cell.

Figure 4 shows an example of the selective eli-sion transform. The transform’s grouping functionselects all table cells whose width is larger than agiven threshold (Figure 4a). When table cells arenested, it examines the inner cells first.

As with the generalized outlining transform,the selective elision transform’s summarizingfunction uses shrinking factor and informationdensity to decide which subgroups to elide. It cal-


Mobile Web Access

Existing Transcoding Heuristics

The existing literature describes seventranscoding heuristics.1,2

• Bickmore and colleagues proposedthe outlining transform1 for paragraphsthat begin with section headers.Theoutlining transform replaces thesection headers with hyperlinkspointing to the corresponding textblocks. This transform effectivelypreserves a Web page’s itemizedstructure while significantly reducingthe display size.

• The improved outlining transform2 adds amore general grouping function andsupports more syntactic combinationsthan the original outlining transform.Transcoding system developers can usethe improved outlining transformwhere conceptually higher (moreabstract) and lower (more detailed)pairs exist.

• Transcoding system developers usethe first sentence elision transform1

when a Web page’s text blocks are toolarge to be displayed on a handhelddevice. A hyperlink hides all but thefirst sentence of the text block.Thistransform works well when the firstsentence summarizes the entire block.

• The restricted first sentence elision

transform2 has a more limited groupingfunction than the first sentence elisiontransform. If a long text block is withina table structure, or a text blockincludes a table structure, therestricted transform suppresses thefirst sentence elision transform tomaintain the Web page’s tablestructure. By not applying the firstsentence elision transform where tablestructures exist, this transform letstable-specific heuristics handle tables.

• The image reduction and elisiontransforms1 are useful in dealing withimages in Web pages.They scale downimages with a predefined scaling factorand create hyperlinks pointing to thereduced images.

• The indexed segmentation transform1

divides a long Web page into asequence of small subpages that fit ahandheld device’s display. Thetransform tries to find logicalelements, such as text blocks or lists,by analyzing syntactic information onthe Web page. It sequentially arrangesthe identified elements in thetranscoded page until it can properlydisplay the new page on the handhelddevice.The transform then creates asequence of subpages, each connected

via hyperlinks.• The table transform1 identifies a table in

a Web page and checks that a handhelddevice can display it properly. If thetable is too wide or too long, thetransform unrolls it and creates onesubpage per cell in a top-down, left-right order.

Existing transcoding heuristics employgrouping functions that are effective onlyfor small combinations of HTML tags. Fur-thermore, because the grouping functionsconsider only the local syntactic attributeswithout weighing a Web page’s overall lay-out, they cannot properly reflect the devel-oper’s intention, especially for complexWeb pages.Our structure-aware transcod-ing heuristics improve on these by consid-ering both the importance of Web compo-nents and the Web page structure whenmaking transcoding decisions.

References1. T.Bickmore, A.Girgensohn, and J.W.Sullivan,“Web

Page Filtering and Reauthoring for Mobile Users,”

Computer J., vol. 42, no. 6, 1999, pp. 534–546.

2. Y. Hwang et al.,“WebAlchemist: A Web Transcod-

ing System for Mobile Web Access in Handheld

Devices,” Proc. ITCom, SPIE — The Int’l Soc. for

Optical Eng., 2001, pp. 37–46.

culates these parameters similarly to the general-ized outlining transform.

The center cell in Figure 4a has higher infor-mation density because it is wider than the others,and is therefore not elided. As Figure 4b shows,the selective elision transform preserves the impor-tant table structure of the original CNN homepage.

Experiments To evaluate the effectiveness of the proposedtranscoding heuristics, we extended the prototypeof the WebAchemist transcoding system, an in-house testbed for Web transcoding research.8 Weimplemented our new heuristics and modifiedWebAlchemist’s transcoding manager module toreflect the new transcoding parameters, informa-tion density and shrinking factor.

Transcoding Manager The extended WebAlchemist system supports sixtranscoding heuristics: the image reduction andelision transforms, the restricted first sentence eli-sion transform, the indexed segmentation trans-form, the improved outlining transform, the gen-eralized outlining transform, and the selectiveelision transform. We also modified the first fourtranscoding heuristics to work with informationdensity and the shrinking factor.

The transcoding manager’s main role is todecide how to use the six heuristics. Becausetranscoding heuristics require several parametersfor proper operation (the display size of a hand-held device and predefined threshold values, forexample), the transcoding manager also decideswhich value to use for each parameter. We empir-ically determined the following sequence of steps

as a default sequence for WebAlchemist:

1. improved outlining transform2. generalized outlining transform3. selective elision transform4. restricted first sentence elision transform5. image reduction and elision transforms6. indexed segmentation transform

This sequence follows our intuition that preserv-ing the overall layout (or structure) is an impor-tant requirement for high-quality transcoding ofcomplex Web pages. The first three transformseffectively reduce display size while keeping theoriginal page structure.

Subjective Evaluation To evaluate WebAlchemist’s effectiveness in con-verting complex Web pages, we chose a subjectiveevaluation. Initial experiments showed that similarobjective characteristics can exhibit striking dif-ferences when evaluated subjectively.

Forty-three college students and engineers liv-ing in Seoul participated in the evaluation. Weasked each participant, all of whom are active Webusers, to judge the transcoded pages’ quality usingone of five grades: fair, good, and excellent are sat-isfactory ratings while unusable and poor areunsatisfactory. We used 13 Web sites as testpages,7 four of which were well-known news sites(for example, www.cnn.com) and four were popu-lar portals/search engines (such as www.yahoo.com). A simulated 320 × 240 display (typical forPDAs) showed both transcoded and original pages.

Figure 5 summarizes the results of our subjec-tive evaluation. Figure 6 shows a transcoded ver-


Web Transcoding

Figure 4. An example of the selective elision transform. (a) The CNN homepage and (b) the transcodedCNN homepage after applying the selective elision transform.

(a) (b)

sion of the Yahoo homepage. (All transcoded Webpages are available at http://davinci.snu.ac.kr/WebAlchemist/experiments/.)

As Figure 5 shows, at least 80 percent of the eval-uators gave the test pages satisfactory ratings. Fourof the pages — Lab_member, Mobicom, Nasa.gov,and Gnu.org — received excellent ratings from morethan 80 percent of the evaluators. Furthermore, noevaluator rated any test Web page unusable.

One interesting observation from this evalua-tion is that most mobile users avoid horizontal

scrolls as much as possible, a fact that contributedto the CNN homepage’s relatively poor rating.Considering the complexity of the tested Webpages, the evaluation results demonstrate WebAl-chemist’s overall effectiveness.

ConclusionAlthough the current version of WebAlchemistalready produces usable Web pages for handhelddevices, we can further improve it within the auto-matic reauthoring framework. For example, know-ing a Web component’s access profile can helpdetermine its relative importance. We plan toextend WebAlchemist by incorporating user accessprofile into the existing transcoding heuristics.

AcknowledgmentsA Korea Research Foundation grant (KRF-2001-041-E00243)

supported this work.

References

1. K.H. Britton et al., “Transcoding: Extending E-Business to

New Environments,” IBM Systems J., vol. 40, no. 1, 2001,

pp. 153-178.

2. E. Kaasinen et al., “Two Approaches to Bringing Internet

Services to WAP Devices,” Proc. 9th WWW Conf., ACM

Press, 2000, pp. 231-246.

3. J. Freire, B. Kumar, and D.F. Lieuwen, “Webviews: Access-

ing Personalized Web Content and Services,” Proc. 10th

WWW Conf., ACM Press, 2001, pp. 576-586.

4. T. Bickmore, A. Girgensohn, and J.W. Sullivan, “Web Page

Filtering and Reauthoring for Mobile Users,” Computer J.,


Mobile Web Access

Figure 5. Subjective evaluation results. At least 80 percent of theparticipants rated the test pages as satisfactory or better.

Was

hing

ton

Post

CN

N

LA T

imes

Cio

sin

Empa

s

Hot

bot

Lyco

s

Yaho

o

Yaho

o_re

s

Lab

mem

ber

Mob

icom

NA

SA

Gnu

100

90

80

70

60

50

40

30

20

10

0

Perc

ent

UnusablePoorFairGoodExcellent

Figure 6. Transcoding example. (a) Original Yahoo.com homepage and (b) WebAlchemist-transcoded version of the page.

(a) (b)

vol. 42, no. 6, 1999, pp. 534-546.

5. T. Bickmore and W. Schilit, “Digestor: Device-Independent

Access to the World Wide Web,” Computer Networks and

ISDN Systems, vol. 29, no. 8, 1997, pp. 1075-1082.

6. J. Chen et al., “Function-Based Object Model Towards Web-

site Adaptation,” Proc. 10th WWW Conf., ACM Press, 2001,

pp. 587-596.

7. Y. Hwang, J. Kim, and E. Seo, Structure-Based Transcod-

ing Heuristics for Mobile Devices, tech. report, School of

Computer Science and Eng., Seoul National Univ., 2002;

http://davinci.snu.ac.kr/Download/hwang_techrep.pdf.

8. Y. Hwang et al., “WebAlchemist: A Web Transcoding System

for Mobile Web Access in Handheld Devices,” Proc. ITCom,

SPIE — The Int’l Soc. for Optical Eng., 2001, pp. 37-46.

Yonghyun Hwang is a full-time instructor at the Naval Acade-

my, Korea. His research interests include Web transcoding

for mobile systems, mobile communications, and embed-

ded systems. Hwang received an MS in computer science

and engineering from Seoul National University, Korea.

Contact him at [email protected].

Jihong Kim is an associate professor in the School of Comput-

er Science and Engineering, Seoul National University,

Korea. His research interests include embedded systems,

computer architecture, real-time systems, and Java com-

puting. Kim received a PhD in computer science and engi-

neering from the University of Washington. For questions

and comments about this article, contact him at School of

Computer Science and Engineering, San 56-1 Shilim-

Dong, Kwanak-Ku, Seoul National University, Seoul, Korea

151-742; [email protected].

Eunkyong Seo is a senior student in the School of Computer

Science and Engineering, Seoul National University, Korea.

Her research interests include optimizations, algorithm

design and analysis, and embedded systems. Contact her

at [email protected].


Web Transcoding

EXECUTIVE STAFFExecutive Director: DAVID W.HENNAGEAssoc. Executive Director:ANNE MARIE KELLYPublisher: ANGELA BURGESSAssistant Publisher: DICK PRICEDirector, Administration: VIOLET S. DOANDirector, Information Technology & Services: ROBERT CAREManager, Research & Planning: JOHN C.KEATON

COMPUTER SOCIETY OFFICESHeadquarters Office1730 Massachusetts Ave. NW

Washington, DC 20036-1992

Phone: +1 202 371 0101 • Fax: +1 202 728 9614

E-mail: [email protected]

Publications Office10662 Los Vaqueros Cir., PO Box 3014

Los Alamitos, CA 90720-1314

Phone:+1 714 8218380


Membership and Publication Orders:

Phone: +1 800 272 6657 Fax: +1 714 821 4641


Asia/Pacific OfficeWatanabe Building

1-4-2 Minami-Aoyama,Minato-ku,

Tokyo107-0062, Japan

Phone: +81 3 3408 3118 • Fax: +81 3 3408 3553


PURPOSE The IEEE Computer Society is theworld’s largest association of computing profes-sionals, and is the leading provider of technicalinformation in the field.

MEMBERSHIP Members receive the month-ly magazine COMPUTER, discounts, and oppor-tunities to serve (all activities are led by volun-teer members). Membership is open to all IEEEmembers, affiliate society members, and othersinterested in the computer field.

BOARD OF GOVERNORSTerm Expiring 2003: Fiorenza C. Albert-Howard, Manfred Broy, Alan Clements, Richard A.Kemmerer, Susan A. Mengel, James W. Moore,Christina M. SchoberTerm Expiring 2004: Jean M. Bacon, RicardoBaeza-Yates, Deborah M. Cooper, George V. Cybenko,Haruhisha Ichikawa, Lowell G. Johnson, Thomas W.WilliamsTerm Expiring 2005: Oscar N. Garcia, Mark AGrant, Michel Israel, Stephen B. Seidman, KathleenM. Swigger, Makoto Takizawa, Michael R. Williams

Next Board Meeting: 22 Nov. 2003, Tampa, FL

IEEE OFFICERSPresident: MICHAEL S. ADLERPresident-Elect: ARTHUR W. WINSTONPast President: RAYMOND D. FINDLAYExecutive Director: DANIEL J. SENESESecretary: LEVENT ONURALTreasurer: PEDRO A. RAYVP, Educational Activities: JAMES M. TIENVP, Publications Activities:MICHAEL R. LIGHTNERVP, Regional Activities: W. CLEON ANDERSONVP, Standards Association: GERALD H. PETERSONVP, Technical Activities: RALPH W. WYNDRUM JR.IEEE Division VIII Director JAMES D. ISAAKPresident, IEEE-USA: JAMES V. LEONARD

EXECUTIVE COMMITTEEPresident:STEPHEN L. DIAMOND* Picosoft, Inc.P.O.Box 5032San Mateo, CA 94402Phone: +1 650 570 6060Fax: +1 650 345 [email protected]

President-Elect: CARL K. CHANG*Past President: WILLIS. K. KING*VP, Educational Activities: DEBORAH K. SCHERRER(1ST VP)*VP, Conferences and Tutorials: CHRISTINASCHOBER*VP, Chapters Activities: MURALI VARANASI†VP, Publications: RANGACHAR KASTURI †VP, Standards Activities: JAMES W. MOORE†VP, Technical Activities: YERVANT ZORIAN†Secretary: OSCAR N. GARCIA*Treasurer:WOLFGANG K. GILOI* (2ND VP)2002–2003 IEEE Division VIII Director: JAMES D.ISAAK†2003–2004 IEEE Division V Director: GUYLAINE M.POLLOCK†2003 IEEE Division V Director-Elect: GENE H.HOFFNAGLEComputer Editor in Chief: DORIS L. CARVER†Executive Director: DAVID W. HENNAGE†

* voting member of the Board of Governors† nonvoting member of the Board of Governors

COMPUTER SOCIETY WEB SITEThe IEEE Computer Society’s Web site, athttp://computer.org, offers information andsamples from the society’s publications and con-ferences, as well as a broad range of informationabout technical committees, standards, studentactivities, and more.

Structure-aware web transcoding for mobile …web.eecs.umich.edu/~zmao/eecs589/papers/transcode.pdfStructure-Aware Web Transcoding for Mobile Devices Transcoding heuristics that consider

Documents