Data Mountain:
`Using Spatial Memory for Document Management
`
`George Robertson, Mary Czerwinski, Kevin Larson,
`Daniel C. Robbins, David Thiel, and Maarten van Dantzich
`Microsoft Research
`One Microsoft Way
`Redmond, WA 98052,USA
`Tel: 1-425-703-1527
`E-mail: ggr@microsoft.com
`
`ABSTRACT
`
`Effective management of documents on computers has
`been a central user interface problem for many years.
`One common approach involves using 2D spatial layouts
`of icons representing the documents, particularly for in-
`formation workspace tasks. This approach takes advan-
`tage of human 2D spatial cognition. More recently, sev-
`eral 3D spatial layouts have engaged 3D spatial cognition
`capabilities. Some have attempted to use spatial memory
`in 3D virtual environments. However, there has been no
`proof to date that spatial memory works the same way in
`3D virtual environments as it does in the real world. We
`describe a new technique for document management
`called the Data Mountain, which allows users to place
`documents at arbitrary positions on an inclined plane in a
`3D desktop virtual environment using a simple 2D inter-
`action technique. We discuss how the design evolved in
`response to user feedback. We also describe a user study
`that shows that the Data Mountain does take advantage of
`spatial memory. Our study shows that the Data Mountain
`has statistically reliable advantages over the Microsoft
`Internet Explorer Favorites mechanism for managing
`documents of interest in an information workspace.
`
`KEYWORDS
`
`3D user interfaces, desktop VR, information visualization,
`spatial cognition, spatial memory, document management
`
`INTRODUCTION
`
`Managing documents effectively on computers has been a
`key user interface design problem for the last thirty years.
`The issue has become more critical as users venture onto
`the World-Wide Web, because the number of easily ac-
`cessible documents has increased dramatically. Graphics
`technology, processor speed, and primary memory capac-
`ity advances have made it possible to build systems that
`
`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 the first page. To copy otherwise, to
`republish, to post on servers or to redistribute to lists, requires prior specific
`permission and/or a fee.
`UIST ’98. San Francisco, CA
`(cid:211) 1998 ACM 0-58113-034-1/98/11... $5.00
`
`Figure 1: Data Mountain with 100 web pages.
`
`help with this document management problem.
`
`The Data Mountain (Figure 1) is a novel user interface for
`document management designed specifically to take ad-
`vantage of human spatial memory (i.e., the ability to re-
`member where you put something). In our current proto-
`type, the user freely arranges document thumbnails on an
`inclined plane textured with passive landmarks. We use
`3D visual and audio cues to enhance the similarity to real-
`world object arrangement, yet use simple 2D interaction
`techniques and common pointing devices (like the mouse)
`for all interactions. The system is designed with a fixed
`viewpoint, so users need not navigate around the space.
`Users can identify and distinguish documents both
`through their thumbnail representation and through pop-
`up titles.
`
`In this paper we describe the document management task
`and discuss existing graphical solutions. We then discuss
`related work from the field of spatial cognition, and issues
`of navigation and document management in the specific
`context of the World-Wide Web. Next, we describe the
`Data Mountain in detail, and report on a user study that
`compared it to the Microsoft Internet Explorer (IE4) Fa-
`vorites mechanism. The paper concludes with a discus-
`sion of study findings and planned future work.
`
`153
`
`APPLE 1030
`
`1
`
`

`

`Document Management
`Document management tasks occur in a variety of con-
`texts, and over a wide range of sizes of information stores
`and information structures. For example, tasks include
`managing files in a file system, mail messages, and web
`pages. The basic information structures a user might en-
`counter include unordered sets, ordered lists, hierarchies,
`and graphs. Often the documents may belong to more
`than one of these information structures. Document
`browsing, searching, overviews, histories, and informa-
`tion workspaces all employ such structures.
`
`The concept of information workspaces, introduced by
`Card, Robertson, and Mackinlay in 1991 [4], refers to the
`environment in which documents of interest can be com-
`pared, manipulated, and stored for future use. Iconic
`desktops, web browser Favorites or Bookmarks, and the
`Web Forager [5] are examples of information workspaces.
`
`RELATED WORK
`
`Document Management Systems
`Early graphical methods for document management in-
`cluded list views, expandable lists for viewing hierar-
`chies, and iconic 2D spatial layouts. The Apple Macin-
`tosh (circa 1984) included list views and a spatial layout
`(icon view). The spatial layout allowed the user to place
`icons in whatever grouping the user desired. Apple later
`added expandable lists for hierarchies, and piles [18].
`Piles enrich the spatial layout by allowing the user to
`group related documents and take less screen space for
`the group. Yet another form of 2D spatial layout is the
`Treemap [14], which is a space-filling layout that is gen-
`erated automatically, used primarily for overviews of
`document collections and their meta-data.
`
`SemNet [10] was an early 3D spatial layout of documents.
`It tackled the difficult problem of visualizing networks.
`The result was difficult to understand because of the
`complexity of the information and layout. The Informa-
`tion Visualizer project [4][22] at Xerox PARC introduced
`a broad set of 3D visualization and interaction techniques
`for understanding information. In 1994, Maya Design
`Group introduced Workscape [3] as the first example of a
`3D spatial layout of documents under the user’s control.
`The Web Forager [5] built on the experience gained from
`the Information Visualizer project and introduced a 3D
`spatial layout for web pages and WebBooks [5].
`
`The Visible Language Workshop of the MIT Media Lab
`did research in the design of dynamic virtual information
`spaces, combining typography and 3D graphics layout to
`present visually appealing interactive information land-
`scapes [16][21].
`
`Typical 2D desktops (Windows or Mac OS) use a spatial
`layout of icons and overlapping windows. As will be
`seen, the Data Mountain supports a larger number of ob-
`jects, prevents overlap with a page avoidance algorithm,
`and uses thumbnail images instead of icons.
`
`Figure 2: Selected page in preferred viewing position.
`
`PadPrints [13] uses the Pad++ 2D zoomable user interface
`to implement a thumbnail image-based web history
`mechanism that is superior to text based history mecha-
`nisms. The authors raise a question about whether their
`technique is successful because of the thumbnail images
`or because of the zoomable interface. PadPrints uses an
`automatic layout for short term use; the Data Mountain
`uses a manual layout to exploit spatial memory for long
`term use.
`
`Many of the techniques mentioned made use of spatial
`cognition, whether or not this was done intentionally. In
`particular, automatic spatial layouts of information lever-
`age the user’s ability to recognize and understand spatial
`relationships (both in 2D and 3D). The 3D interfaces
`make it possible to display more information without in-
`curring additional cognitive load, because of pre-attentive
`processing of perspective views (i.e., smaller size indi-
`cates spatial relationships at a distance). The Maya De-
`sign Group Workscape system and the Web Forager in-
`tended to use spatial memory, by allowing the user to
`place documents as an aid to finding them again.
`
`The Data Mountain is an advance over Workscape and
`Web Forager in several ways. First, the Data Mountain
`allows the user to place the document at an arbitrary loca-
`tion on its slope with a simpler interaction technique than
`the earlier systems (taking advantage of constrained mo-
`tion along the inclined plane of the Mountain). Second,
`when a page is being moved, other pages are moved out
`of the way (active page avoidance), yet the user still sees
`visual cues indicating where every page will be when the
`movement is completed. Third, the Data Mountain ex-
`ploits a variety of audio cues to augment the visual cues.
`Fourth, page titles are displayed whenever the mouse
`moves over a page. And fifth, in light of research in spa-
`tial cognition and wayfinding, visual neighborhood de-
`marcation cues are provided to assist the user in arranging
`her personal space on the Data Mountain.
`
`154
`
`2
`
`

`

`Figure 3: Second subject’s layout of same 100 pages.
`
`Figure 4: Third subject’s layout of same 100 pages.
`
`Spatial Cognition
`There is a large body of literature on spatial cognition (see
`[23][7] for recent examples) and wayfinding [8][9][25],
`both for real and electronic worlds. Some of these studies
`have culminated in a set of guidelines for designers of
`virtual worlds. For instance, leveraging knowledge from
`the architectural domain [17][19], Darken and Silbert [9]
`have shown that adding real world landmarks, like bor-
`ders, paths, boundaries and directional cues, can greatly
`benefit navigation performance in virtual reality. In their
`studies, Darken and Silbert have shown that stationary or
`predictably moving cues are optimal, and that multiple
`sensory modalities can be combined to assist searching
`through an electronic space (like 3D sound cues). They
`also have shown that if the space is not divided using a
`simple, organizing principle, users will impose their own,
`conceptual organization upon the space.
`
`DATA MOUNTAIN
`
`The Data Mountain is a 3D document management sys-
`tem. The current prototype is being used as an alternative
`to current web browser Favorites or Bookmark mecha-
`nisms, so we sometimes refer to pages as the objects that
`appear on a mountain. It should be understood that other
`forms of documents should work equally well.
`
`When a page is first encountered, it appears in a preferred
`viewing position (see Figure 2), so that it is easily read.
`The user can place the page by dragging it with a tradi-
`tional left-mouse-button drag technique. As the page is
`being dragged, other pages move out of the way so the
`page being moved is not occluded. After the page has
`been placed, it can be selected with a single click to bring
`it back to the preferred viewing position. Visual and
`audio cues, as well as the interaction techniques are more
`fully described below.
`
`The Data Mountain is designed to work in a desktop 3D
`graphics virtual environment (also known as desktop VR),
`
`although it could certainly work in either Fishtank VR
`[27] or VR with head-tracked head-mounted displays.
`Examples described here are desktop VR examples.
`
`Leveraging Natural Human Capabilities
`The primary motivation for the design of the Data Moun-
`tain came from a desire to leverage natural human capa-
`bilities, particularly cognitive and perceptual skills. In
`particular, 3D perception is used to allow for the repre-
`sentation of a large number of web page thumbnails with
`minimal cognitive load. Our pre-attentive ability to rec-
`ognize spatial relationships based on simple 3D depth
`cues (like perspective views and occlusion) makes it pos-
`sible to place pages at a distance (thereby using less
`screen space) and understand their spatial relationships
`without thinking about it. We can leverage audio percep-
`tion to reinforce what is happening in the visual channel.
`Both the visual and auditory perception can enable basic
`human pattern recognition capabilities. And finally, we
`hope to use spatial memory to make it easier to find
`documents in the information workspace.
`
`In the real world, spatial memory often aids us in finding
`things. For example, when we place a piece of paper on a
`pile in our office, we are likely to remember approxi-
`mately where that paper is for a long time. Our hope is
`that this ability works as well in a virtual space as it does
`in the physical world. This is not an obvious conclusion.
`However, if spatial memory is primarily an act of building
`a mental map of the space, then we should be able to do
`the same thing in a virtual environment, and take advan-
`tage of it.
`
`Data Mountain Visual and Audio Design
`The Data Mountain provides a continuous surface on
`which documents are dragged. The document being
`dragged remains visible so the user is always aware of the
`surrounding pages. This is in direct contrast to the way a
`Web Forager [5] user places documents in a discrete set
`of tiered locations using flicking gestures. We believe the
`
`155
`
`3
`
`

`

`Data Mountain Interaction Design
`When the user clicks on a page stored on the Data Moun-
`tain, the page is moved forward to a preferred viewing
`position, as shown in Figure 2. The animation to bring
`the page forward lasts about one second [4], uses a slow-
`in/slow-out animation [6][11], and is accompanied by an
`audio cue. We use a higher resolution texture map for the
`page image in the preferred viewing position, ensuring
`that the page is quite readable.
`
`When in the preferred viewing position, a click on the
`page will either select and follow a hyperlink, or put the
`page back on the Data Mountain in its last known loca-
`tion. This is also done with a one second, slow-in/slow-
`out animation accompanied by audio.
`
`We provided a pop-up label similar to tool-tips to display
`page titles. Subjects tended to use their spatial memory to
`get to the neighborhood of the page, then riffle through
`the titles (like riffling through a pile of papers on your
`desk) to find the page. A standard tool-tip uses a hover
`time before the tip is displayed. We determined in a pilot
`study that the hover time was not effective since it pre-
`cluded rapid inspection of multiple titles. Hence, the title
`appears as soon as the mouse moves over a page. An
`example is shown in Figure 5. In one group from our user
`study, the title was shown just above (but disconnected
`from) the page. Some subjects could not easily distinguish
`which thumbnail the title applied to, so our second Data
`Mountain design added an identically colored halo around
`the thumbnail, creating a visual link to the title.
`
`A page can be moved at any time by dragging it with the
`mouse. Since the page is visible during the move, the
`user knows where the page will be when the drag is ter-
`minated. The movement is continuous and constrained to
`the surface of the Data Mountain. This results in one of
`the principal advantages of the Data Mountain; the user
`gets the advantages of a 3D environment (better use of
`space, spatial relationships perceived at low cognitive
`overhead, etc.), but interacts with it using a simple 2D
`interaction technique.
`Page Avoidance Behavior
`When moving a page, what is the right behavior for pages
`that are encountered (i.e., how are collisions handled)?
`We have tried three alternatives, each improvement
`driven by user comments. First, we did nothing. That is,
`the page in motion simply passed through the pages in the
`way. This approach suffers from several problems. It
`makes the system seem lifeless and makes the metaphor
`harder to understand. In addition, it is quite easy to put
`pages right on top of each other, making it difficult to find
`some pages.
`
`Second, we tried a simulation of tall grass. Think about
`what happens when you walk past tall grass. If you walk
`slowly, the grass moves out of your way slowly then re-
`turns. If you walk fast, the grass seems to fly out of your
`way. We implemented a simple simulation in which pre-
`viously placed pages behave like grass displaced with a
`
`Figure 5: Title shown while hovering over page.
`
`user’s act of directly placing the page on the continuous
`surface of the Data Mountain aids spatial memory.
`
`The Data Mountain prototype uses a planar surface (a
`plane tilted at 65 degrees), as shown in Figure 1. The
`landscape texture on the Data Mountain surface provides
`passive landmarks for the user meant as an aid for group-
`ing objects into categories, but the landmarks have no
`explicit meaning. The user can place the web pages any-
`where on the mountain. In practice, users create meaning
`by organizing the space. In our study, there were many
`ways to lay out the same set of pages (compare Figures 1,
`3, and 4). There is no “right” layout; rather, the layout is
`very personal, has meaning for the individual who created
`it, and evolves over time under user control.
`
`Note that the current prototype provides no mechanism
`for labeling groups of pages with category titles. While
`some users requested this feature, we found they built a
`very accurate mental map of their categories even without
`explicit labels. Some users employed particularly salient
`thumbnails as visual identifiers of their groups, keeping
`them in front of all other members of the group (thus cre-
`ating their own landmarks).
`
`There are a number of 3D depth cues designed to facili-
`tate spatial cognition. The most obvious are the perspec-
`tive view and occlusion, particularly when pages are be-
`ing moved. The landmarks also offer an obvious cue,
`which may or may not be utilized during page placement
`as well as retrieval. Less obvious, but also quite impor-
`tant, are the shadows cast by the web pages.
`
`Subtle but pervasive audio cues accompany all animations
`and user actions to reinforce the visual cues. The sound
`effects are highly dynamic. For example while moving a
`page the user hears a humming sound that changes pitch
`based on the speed of the page as it is dragged, as well as
`indicating spatial location by controlling volume, low
`pass filtering, panning, and reverb level. As the user
`moves a page, other pages move out of the way as
`needed, producing another distinctive sound.
`
`156
`
`4
`
`

`

`page the user is dragging. Each page that is moved out of
`the way uses a ½ second animation to move aside, fol-
`lowed by a one second animation to move back. This
`feels very lively but suffers from two problems. It still
`does not solve the problem that two pages can end up in
`the same location: the user may drop a page close to an-
`other document's real location while that document is
`temporarily displaced, soon to return, causing one to oc-
`clude the other. The collision avoidance behavior should
`be designed to eliminate such surprises. Also, the move-
`ment is based on first encounter. In other words, if after
`triggering avoidance the dragged page slows down or
`hovers in-place, the return animation will still take place,
`causing the objects to intersect anyway. Essentially, the
`'grassy technique' works well for continuous dragging but
`tends to be annoying when you slow down. Note that
`slowing down is an essential part of object placement:
`effectively, the grassy technique caused some users to
`create unnecessary occlusion since they would pick the
`just vacated spot of a displaced page for the new page. It
`could be modified to be dependent on your speed; but
`since it is basically an estimate, it will fail in some cases.
`The first group of Data Mountain users described in the
`experiment below used this page avoidance mechanism.
`These problems provided strong motivation for finding
`another method and running another set of subjects, as
`some users effectively lost many pages due to occlusion.
`
`In our current implementation, we continually maintain a
`minimum distance between all pages, even while a page is
`being moved, and transitively propagate displacement to
`neighbors as necessary. This has the advantage that the
`user dragging the page continually sees what state will
`result when the drag is terminated (i.e., there is no anima-
`tion settling time). Also, the pages never get fully ob-
`scured. In particular, you cannot move a page and leave
`two pages in the same location. On the other hand, dis-
`placements may propagate far afield when a cluster of
`closely packed pages is 'pushed' by a dragged page, re-
`sulting in more visual unrest than is really desirable. Still,
`the approach feels quite lively, and was used by the final
`group of subjects in our study. We feel this change con-
`tributed most to improved user performance in the second
`Data Mountain group.
`
`Implementation
`The Data Mountain prototype runs under Windows NT
`version 4 on PCs equipped with Intergraph Intense 3D
`Pro 1000 or Pro 2200 graphics accelerators. All applica-
`tion code was written in C++, utilizing our own libraries
`for animation and scene-graph management. These li-
`braries in turn used the ReActor [2] infrastructure and
`OpenGL as the underlying graphics library.
`
`The interactive sound for the Data Mountain was based
`on MISS (the Microsoft Interactive Sound Sequencer)
`which takes parametric sound events and sequences them
`using MIDI to communicate to the wavetable synthesizer
`on a Creative Labs AWE64 Gold card.
`
`For prototyping and user study purposes, the web pages in
`the current implementation are not “live”, i.e.; one cannot
`select and follow a hyperlink. Future versions of the Data
`Mountain will contain live web browsing capabilities.
`
`The 100 pages used in the study below are screen snap-
`shots of actual web pages in 24-bit color. We employ two
`bitmap sizes of each page for texture mapping: a small
`64x64 pixel version (12KB each) for the thumbnails on
`the Data Mountain surface, and a 512x512 pixel version
`(768 KB each) for the close-up view. One hundred
`thumbnails plus a close-up together will fit in 2MB of
`texture cache. Our system implements text labels using
`texture-mapped fonts [12]. This is vastly preferable to
`vector fonts, and has the advantage that it enables display
`of legible text on surfaces that are not screen-aligned.
`Category labels placed on the Data Mountain surface are
`just one potential use of such perspective-distorted text.
`
`If pop-up labels are naively attached to thumbnails, they
`will be subject to perspective projection, and thus be
`smaller for pages that are placed towards the back of the
`information workspace. We found it difficult to choose an
`absolute label size in model coordinates that produces
`appropriately scaled text for labels of both foreground and
`background documents. Instead, we implemented labels
`for thumbnails to be of identical size, independent of the
`document's distance from the viewer in virtual space. We
`do this by placing the title-tip a constant distance away
`from the eye-point, on the vector from the eye-point to the
`page whose title we are showing.
`
`USER STUDY
`
`Studies by Tauscher and Greenberg [24] and Abrams [1]
`were the earliest attempts to gather information about user
`behavior as they traverse the web over several months.
`According to Abrams [1], users develop their own per-
`sonal web information spaces through the use of Favorites
`mechanisms in order to combat the problems of informa-
`tion overload, pollution, entropy, structure and lack of a
`global view of the web. Users do this by building a
`smaller, more valuable, organized and personal view of
`the web.
`
`Usage tracking shows that hotlists, bookmarks and Fa-
`vorites folders are the navigation tools most frequently
`utilized by users for locating information on the web [20].
`
`Hence, web browser designers need to provide their users
`with mechanisms for creating personal web information
`spaces that can reliably and efficiently return the user to
`their favorite web sites. Implementing such mechanisms
`relaxes the cognitive and temporal demands of hypertext
`navigation [1]. Usability studies, as well as basic research,
`however, indicate that the current designs for navigating
`the web are still sub-optimal in supporting users’ cogni-
`tive models of web spaces and the amount of information
`they need to repeatedly consume [1][24].
`
`The Abrams study [1], in particular, pointed out how epi-
`sodic memory [26], or memory for events, could be
`
`157
`
`5
`
`

`

`thought of as a primary cognitive avenue for retrieving
`web pages from Favorites. As pointed out in that re-
`search, reviewing a set of bookmarks or Favorites is basi-
`cally a process of using textual cues to retrieve a memory
`for how that web page was stored. Usage data from that
`study showed that users had trouble retrieving their fa-
`vorite web pages, often because the default title which
`they used to store the web page was an inadequate re-
`trieval cue for recognition. Left unanswered, though, was
`whether and to what extent spatial cognition was playing
`a role in users’ navigational behaviors, and to what extent
`our browser designs could leverage what is known about
`spatial cognition and wayfinding. As users’ personal web
`information spaces grow larger, how can we effectively
`design the Favorites user interface to afford efficient re-
`trieval? One way to increase the amount of information
`on the screen is to move to a 3D user interface. Also, the
`possibility of adding spatial landmarks, edges and audio
`cues, as discussed above, should prove useful. These are
`the issues that are explored in the current set of studies.
`
`The primary user interface design for a Favorites folder is
`similar to the hierarchical tree views used to browse files
`in a computer file system. A user is allowed to enter
`items into an organized list, often alphabetized, and the
`list can have any number of subcategory structures added
`to it. Often, the URL or the web page title is the default
`label representing the web page when browsing the list.
`The list is text only, so does not allow users to leverage
`other channels of information that may also be effective
`when attempting to retrieve web pages, such as the audi-
`tory channel. The question of interest is how effective is
`the Data Mountain for leveraging all aspects of memory
`during the retrieval of a web page in Favorites?
`
`Our hypothesis was that the effectiveness and usability of
`the Data Mountain depends in part on the transfer of real-
`world spatial memory skills to a virtual environment.
`
`Title
`
`Bezerk - The Free On-
`line Entertainment Net-
`work
`
`Title, Summary, &
`Thumbnail
`Bezerk - The Free On-
`line Entertainment Net-
`work
`The new, FREE premier
`online entertainment
`network features YOU
`DON’T KNOW JACK
`the netshow, the online
`version of the irreverent
`quiz show party game
`CD-ROM!
`
`This assumption was also made for earlier systems that
`make use of spatial memory for document management
`(like the Maya Designs Workscape and the Xerox PARC
`Web Forager). To test this hypothesis, we did a user
`study comparing use of the Microsoft Internet Explorer
`4.0 (IE4) Favorites and the Data Mountain for the same
`storage and retrieval tasks. IE4 was chosen for this com-
`parison because it is a shipping product with which many
`readers will be familiar.
`
`Methods
`Subjects. Thirty-two experienced IE4 users participated
`in this study. All users had to successfully answer a series
`of screening questions pertaining to web browser and
`Internet knowledge in order to qualify for participation.
`Subject ages ranged from 18 through 50 years old, and all
`had normal or corrected-to-normal vision. The number of
`females and males was balanced.
`
`Equipment. The study was run on high-end Pentium ma-
`chines (P6-266 or P6-300), with at least 128 MB of mem-
`ory, and a 17-inch display. The machines had either an
`Intergraph Intense 3D Pro 1000 or 2200 graphics accel-
`erator card and ran Windows NT4. One hundred web
`pages were used in this study; fifty pages were selected
`randomly from PC Magazine’s list of top web sites and
`fifty pages selected randomly from the Yahoo! database.
`A web server was contained on the local computer for
`each user eliminating network lag. This was done to
`maintain consistent system response times across applica-
`tions, users and time of study, thus eliminating possible
`confounds in the results.
`
`Procedure. Users participated in one of three groups.
`One group of users stored and retrieved web pages using
`the standard IE4 Favorites mechanism. The other two
`groups of users stored and retrieved web pages using the
`Data Mountain. The second Data Mountain group (DM2)
`interacted with a version of the Data Mountain that incor-
`Summary
`Thumbnail
`
`The new, FREE premier
`online entertainment
`network features YOU
`DON’T KNOW JACK
`the netshow, the online
`version of the irreverent
`quiz show party game
`CD-ROM!
`
`Table 1: Examples of the four cueing conditions used in the study
`
`158
`
`6
`
`

`

`Data Mountain v. IE4: Reaction Time
`
`Data Mountain v. IE4: Incorrect Retrievals
`
`IE4
`DM 1
`DM 2
`
`16.00
`
`14.00
`
`12.00
`
`10.00
`
`8.00
`
`6.00
`
`4.00
`
`2.00
`
`0.00
`
`Number of Incorrect Retrievals
`
`IE4
`DM 1
`DM 2
`
`20
`
`18
`
`16
`
`14
`
`12
`
`10
`
`02468
`
`Reaction Time (in seconds)
`
`All
`
`Summary
`
`Thumbnail
`
`Title
`
`All
`
`Summary
`
`Thumbnail
`
`Title
`
`Retrieval Cue
`
`Retrieval Cue
`
`Figure 7: Average number of incorrect pages retrieved
`by cueing condition and application.
`
`the session would entail. The IE4 users could use either
`the Favorites panel or the Organize Favorites dialog box
`for storage and organization. After all pages had been
`stored, participants in each group were given an addi-
`tional opportunity to re-organize their Favorites. Note
`that there are differing interaction constraints on the users
`depending on the application they used. When using IE4,
`users were restricted in terms of the default amount of the
`screen space taken up by the Favorites mechanism. The
`Data Mountain users, on the other hand, used the entire
`screen for managing their web page layout by default.
`However, since Data Mountain users were operating on
`cropped images of web pages, not the “live” pages, they
`were restricted in terms of how much they could learn
`about a web page (e.g., they could not scroll down and
`read the contents of a page for more information like the
`IE4 users could and often did). All study participants were
`experienced IE4 users.
`
`During retrieval, which ensued after a short break, the
`participants were shown one of four different retrieval
`cues and asked to find the related page. The four retrieval
`cueing conditions were: the title of the page, a one or two
`sentence summary of the page’s content, a thumbnail im-
`age of the page, and all three cues simultaneously (called
`the “All” cue). Participants saw 25 trials of each cueing
`condition, for a total of 100 retrievals. All pages presented
`for retrieval were seen in the earlier storage phase. The
`web pages to be stored and the subsequent retrieval cues
`were presented in a random order for each participant.
`Table 1 shows an example of each of the four styles of
`retrieval cues. If a participant could not find the target
`page within two minutes, a “time-out” was enacted and
`the participant was instructed to proceed to the next re-
`trieval task. Page Retrieval was defined as selecting an
`item from the IE4 Favorites to be displayed in the main
`browser pane, or bringing a page forward using the Data
`Mountain. Users were not explicitly discouraged from
`producing incorrect retrievals.
`
`Figure 6: Average retrieval times for each condition.
`
`Data Mountain v. IE4: Failed Attempts
`
`IE4
`DM 1
`DM 2
`
`14
`
`12
`
`10
`
`02468
`
`Number of Failed Attempts
`
`All
`
`Summary
`
`Thumbnail
`
`Title
`
`Retrieval Cues
`
`Figure 8: Average number of failed trials by cue condi-
`tion and application.
`
`porated design changes suggested from the first Data
`Mountain group (DM1). The design changes included
`preventing pages from occluding each other during stor-
`age or manipulation, strengthening the association be-
`tween “hover titles” and their corresponding pages in a
`more intuitive manner, and improving the audio feedback
`(e.g. adding spatialization effects). The discussion section
`(below) will detail to what extent the design changes for
`the second Data Mountain group improved user satisfac-
`tion and performance.
`
`Participants in each group were told that they would be
`storing 100 web pages. They were allowed to create any
`organizational structure they wanted in the IE4 Favorites
`mechanism or the Data Mountain, and were encouraged
`to create a structure that mimicked how they stored fa-
`vorite web pages at home or work. Their web page or-
`ganizations from home were collected for comparison
`purposes. The participants were told they would have to
`use their organization for tasks in the second half of the
`test session, though were not told what the second half of
`
`159
`
`7
`
`

`

`Four main dependent variables were used in this study:
`(1) web page retrieval time; (2) the number of incorrect
`pages selected prior to finding the correct page; (3) the
`number of trials for which the participant failed to retrieve
`the correct page within the two-minute deadline; and (4)
`the participants’ subjective ratings