1

AbstractThis paper shows how previous research intonavigation through urban environments, whichhas emerged from the discipline of urbanplanning, can be adapted to enhance the design ofinformation visualisations. The paper draws onKevin Lynch's seminal work on the legibility ofurban landscapes in order to propose a set ofgeneral techniques which can be applied to thetask of information visualisation. It describes aspecific instantiation of these techniques calledLEADS, a legibility system which post-processesthe output of a range of existing visualisationsystems in order to enhance their legibility. Thepaper provides four examples of the application ofLEADS to different information visualisations.Following this, it discusses experimental work,the conclusions of which provide some tentativesupport for the likely success of this approach.The outcomes of this work are both a recognitionof the important relationship between thedisciplines of urban planning and the design ofinformation visualisations as well as moreconcrete algorithms to be used by the designers ofsuch visualisations.1

1.We would like to gratefully acknowledge the support ofthe UK’s Engineering and Physical Sciences ResearchCouncil (EPSRC) in this work.

1. Introduction

This paper concerns the structure and navigationof information visualisations. More specifically, itprovides some observations on how the design ofsuch visualisations might affect people's ability tolearn to navigate them. These observations arethen translated into computational techniques (i.e.algorithms) for enhancing visualisations. In turn,these techniques are backed up with experimentalresults providing some early indications of theirpotential usefulness.

This work has arisen from the observation thatresearch into the relationship between spatialform and navigation has not only been confined tothe realm of information visualisation; there hasbeen a wealth of previous research on this topic inthe fields of urban planning and architecture. Inmany cases, this work has yielded concrete (!)results in terms of well defined design principlesto support navigation in real-world spatialenvironments. The work described in this paperborrows directly from this previous research andapplies it to the task of information visualisation.We focus in particular on the concept of thelegibility of an urban landscape as initiallydeveloped by Kevin Lynch in the 1960s. Lynch'swork focused on how an individual might develop

The Application of Legibility Techniques toEnhance Information VisualisationsRob Ingram and Steve BenfordDepartment of Computer ScienceThe University of NottinghamNottingham, NG7 2RD,UK

email: { rji, sdb } @cs.nott.ac.ukhttp://www.crg.cs.nott.ac.uk/fax: +44 115 951 4254phone: +44 115 951 4225

2

a so called cognitive map of an urban landscape asa result of repeated exposures to it. Following aseries of experiments involving the navigation ofreal cities, he proposed the existence of five keyfeatures of an urban landscape which, if correctlydesigned, could enhance people's ability to form acognitive map. These were: districts, edges,landmarks, paths and nodes.

This paper reviews Lynch's work and then goeson to develop a series of computationalalgorithms for automatically generating oremphasising these features in a range ofinformation visualisations. It describes theimplementation of these algorithms as a“legibility layer” called LEADS which post-processes the output of various informationvisualisations in order to provide additionalnavigation support. As such, this paper is not somuch concerned with the core design ofvisualisation layouts as it is with how existingvisualisations can be enhanced. Some evaluationexperiments are then described, the results ofwhich provide some initial justification for thesealgorithms and for this approach in general.

2. Goals and relationship toprevious work

Before progressing further, it is first necessaryto clearly state the problem that we are trying toaddress and to say something of its relationship toprevious work on navigation and visualisation.Our work builds on two areas of research:navigation within large-scale information systemsand related to this, the design of graphicalinformation visualisations.

A problem which is often encountered in theuse of large computer based information systemsis that of getting lost. For example, consideringthe area of Hypertext and Hypermedia systems fora moment, Elm and Woods [1] define threecategories of being lost:

• Not knowing where to go next;• Knowing where to go but not how to get there;• Not knowing a current location in relation to an

overall context.I addition other categories might be

considered, such as mistaking the global or localframe of reference, erroneously believing oneknows one’s location or the path to the goal, orsimply being disoriented.

Jacob Nielsen addressed this issue in a numberof related papers [2][3][4] citing cases whereusers were unable to return to given locationswithin a hypertext system. A number of solutionshave been proposed to these kinds of navigationproblems, including guides, tours and variousbacktracking mechanisms. Indeed, variants ofthese have been integrated in the currentgeneration of World Wide Web browsers (e.g.bookmarks and hot-lists). These techniques,however, introduce their own problems; forexample, the passive movement mode in guidedtours may inhibit learning of the system andreduces the hypertext to a linear form andalthough the information provided might help insome tasks a more active approach is preferred.As a result, Nielsen has proposed the use ofoverview diagrams (effectively maps) and fish-eye techniques to provide both local detail andglobal context. This naturally leads us on to thesubject of visualisation.

Information visualisation techniques areintended, at least in part, to overcome some ofthese problems for a range of differentinformation systems. So called “focus andcontext” techniques aim to situate a specific itemof information within a representation of itssurrounding context, thereby addressing the thirdof the above categories of being lost. Perhaps thebest known examples of this approach are thePerspective Wall [5] and Cone Tree [6]visualisations from Xerox PARC.

More recently, researchers have begun toexplore the potential of three dimensional

3

information visualisations which utiliseinteractive 3-D graphics to allow one or moreusers to situate themselves within a visualisationand to move about it or fly over it. Authors suchas Thomas Erikson [7] and Kim Fairchild [8] haveargued that three dimensional visualisationtechniques may help increase the amount ofinformation that people can meaningfully manageand, although the merits of the 3-D approachremain largely unproved, there is clearly agrowing interest in this field. Indeed, there havealready been numerous specific examples of threedimensional visualisations, many of the bestknown ones having been reported at the ACM’sseries of Computer Human Interaction (CHI)conferences.

The work described in the paper builds uponprevious work concerning the navigation ofinformation systems and the construction ofinformation visualisations. Our specific aim is tointroduce a framework for automaticallyenhancing existing information visualisations soas to aid users in more easilylearning to navigatethem. As such, we are not directly concerned withthe design of specific visualisation layouts orindeed with application to specific kinds ofinformation systems. Although the examples thatwe use to illustrate our work are largely basedupon document repositories, the World WideWeb and three dimensional visualisations, ourtechniques are intended to be applicable across awider range of visualisation styles and underlyinginformation systems than these.

At this stage, we must also be careful toprecisely state our interest in navigation. We arenot primarily concerned with the mechanics ofcontrolling the movement of viewpoints andembodiments (although we will briefly touch onthis theme later on); nor are we primarilyconcerned with how people find their way aroundunfamiliar information environments (i.e. onesthat they haven’t seen before). Instead, ourspecific aim is to support people in learning to

navigate visualisations as a consequence ofrepeated exposures to them over a period of time.In other words, we are concerned with how peoplecan be aided in gradually learning the structure ofa graphical space. We anticipate that thetechniques that we propose will be mostlyapplicable to visualisations which are persistent,which evolve relatively slowly and which arerepeatedly visited (a visualisation of a region ofthe World Wide Web would be a good example).

Our approach to this particular navigationproblem has been to apply legibility techniques,developed in the domain of urban planning, to thedomain of information visualisation. Thisapproach has also been recently adopted by otherresearchers. Rudy Darken and John Sibert [9]have considered a number of issues associatedwith navigation in virtual environments. Theyhave evaluated a space which makes use oflegibility features such as landmarks and districts(see section 3) to aid navigation tasks in a flatsimulation of the sea containing ship objects. Indeveloping the BEAD system Matthew Chalmers[10][11] has also considered the use of legibilityfeatures, among other techniques, in theproduction of useable information spaces and useslandmarks, edges and districts in a visualisationsystem based on a landscape metaphor.

This work presented here differs from thisresearch in its development of techniques toautomatically construct legibility features in theabstract spaces produced by a variety of existingor future visualisation systems. One of our mainaims has been to accomplish this withoutrequiring the users of visualisation systems toperform the placement of the features manually.Note also that because this research draws heavilyon city planning literature this paper usesterminology from this area rather than the moretraditional terminology of computer graphics.

In summary then, our aim has been to generateand evaluate techniques for automaticallyenhancing the legibility of information

4

visualisations in order to support users in moreeasily learning to navigate them over a period oftime. The remainder of this paper is structured asfollows. Section 3 reviews previous work onnavigation in urban environments and introducesthe key concept of legibility as proposed byLynch. Section 4 adapts the concept of legibilityto the task of information visualisation anddescribes the set of algorithms and techniqueswhich have been developed for the automaticcreation or enhancement of legibility features.Next, section 5 demonstrates the application ofour algorithms to four existing visualisations: anetwork drawing tool, an interactive searchinginterface to a document repository, a scatter graphdrawing tool and a landscape oriented documentvisualisation. Section 6 describes an initialexperiment to assess the effectiveness of thisimplementation. Finally, section 7 offers by somereflections on specific problems encountered withthis work and outlines possible next steps in itsevolution.

3. The legibility of urbanenvironments

Legibility, in the context of urbanenvironments, is a term which has been used formany years in the discipline of City Planning.Work in this area has focused on the ease withwhich people are able to learn the layout of a cityand then use this knowledge to performwayfinding tasks. In his book “The Image of theCity” [12] Kevin Lynch defines the legibility of acity as:

“-the ease with which its parts may be recognised andcan be organised into a coherent pattern-”

Here, Lynch is referring to the formation of acognitive map within the persons mind [13], astructure which is an internal representation of anenvironment which its inhabitants use as areference when navigating to a destination. It is

proposed that cognitive maps fall into two types.Linear or sequential maps are based on movementthrough the space, or sequential images of a route,and what is observed on the journey. Spatial mapsdo not require this reference movement throughthe imagined space and so areas within the map donot need to be linked to routes through the area. Apersons cognitive map may change over time.Typically, when we are new to an area ourcognitive map will be linear but this will usuallyevolve to become increasingly spatial. Also,active exploration of an environment, as opposedto being guided thought it, encourages theformation of richer maps which are more likely tobe spatial. For more detail on the formation ofspatial knowledge Piaget and Inhelder [14] andDowns and Stea [15] are good starting points.

The Image of the City describes experimentscarried out in a number of major US cities whichsuggest how cognitive maps are built up overtime. The experiments involved obtaininginformation from long term inhabitants of thecities in the form of, for example, interviews,written descriptions of journeys through the cityand drawn maps. By examining this data Lynchidentified five major elements of urbanlandscapes which were identified by theinhabitants and used as the main building blocksof their cognitive maps. These features are:• Landmarks. Static and recognisable objects

which can be used to give a sense of locationand bearing. Examples of landmarks might beprominent buildings or monuments or, on asmaller scale, recognisable shopfronts orroadside installations.

• Districts. Sections of the environment whichhave a distinct character which providescoherence, allowing the whole to be viewed asa single entity. Districts may be identifiable,for example, by the nature of the architectureof the buildings in the area or by their use.

• Paths. Major avenues of travel through theenvironment such as major roads or footpaths.

5

• Nodes. Important points of interest alongpaths, e.g. road junctions or town squares.There is clearly a strong link between paths andnodes.

• Edges. Structures or features providingborders to districts or linear obstacles.Examples might be the waterfront in citiesbuilt on large rivers, or a major road. Note inthe latter case that the road may have a dualnature, being a path for someone travelling in acar but an edge to the pedestrian.Urban environments also contain other cues

for navigation which we might consider includingin virtual worlds. In particular we be interested inthe use of signposts. In city planning there is somecontention over the use of signs as they aresometimes considered to be, to some extent, anadmission of failure in the initial design of anavigable space [16]. However, other researcherspoint out the extent with which signs areintegrated with the everyday process ofwayfinding and so consider them an essentialelement in the urban landscape. We will return tothe subject of signposts later on.

4. Legibility techniques forinformation visualisation

In this section we consider how the generalnotion of legibility and the five specific legibilityfeatures identified by Lynch might be adapted foruse in information visualisation systems. Thealgorithms proposed here were initially presentedin [17]; we represent them here for completenessbefore progressing to discuss recent applicationsand experimental evaluation.

We have stated above that the formation ofcognitive maps is a learning process, so the spacesto which we apply legibility enhancements mustallow learning to happen. This leads us to definethe following criteria for spaces to which outtechniques might most usefully be applied. First,

the space, or the visualisation, ought to persistover a long period of time. This is necessary inorder that enough time is available for learning ofthe space to take place. Second, the information inthe space must be relatively stable whenconsidered in terms of the ratio between thenumber of changes occurring and the overallvolume of information present. Two problemsmay arise if the data is constantly in flux; it maybe difficult to actually produce and place thelegibility features in a useful way; and even if wecould they would quickly become irrelevant,either because they no longer related to the data orbecause they moved too frequently to act as usefulreference points. Finally, the space should beavailable for users to re-enter and use repeatedly.This will allow learning over time to take placenaturally.

The usefulness of legibility features thenobviously depends on the whole data space beingused in a browsing mode rather than viewing onlypieces of data which match a search query ortravelling directly to such objects in the space. Webelieve that a useful aspect of visualisationtechniques is the ability to present data in thecontext of related pieces of information and hencethat browsing is an important part of using such asystem. Chalmers [10] notes that largeinformation spaces are often difficult to accessbecause the user has difficulty recognising aneffective starting point for browsing the data. Hesuggests that searches can provide this ‘way into’the information. It is our hope that by enhancingthe process of learning the layout of the space theuse of legibility features may also help usersestablish an initial starting point for browsing.

The following sub-sections now describe howeach of Lynch’s five legibility features might beintroduced into a general informationvisualisation. This order in which they arepresented (districts, edges, landmarks and thenpaths and nodes) is significant as the placement ofsome of the features depends on others which

6

have already been generated.We also describe how these proposed

techniques have been implemented in a prototypesystem called LEADS (LEgibility for AbstractData Spaces). LEADS has been constructed as anexample “legibility layer”, an independent sub-system which, following some configuration andbasic integration, is able to post-process theoutput of other visualisation systems in order toenhance their legibility. LEADS is referred tothroughout as a concrete example of how ourtechniques can actually be realised. However, weurge the reader to bear in mind that otherimplementations are possible.

Returning briefly to the nature of theinformation visualisation space, most suchsystems can be very broadly classified into twotypes: those which map attributes of the datadirectly onto the axes of the space, in the mannerof a scatter graph, or those which use someincremental algorithm to draw similar objectscloser together. We would like to point out thatthe techniques we are describing here may be, andhave been, applied to both types of space withequal efficacy.

4.1. Districts

The discovery of districts within the data is amatter of finding groups of items which havestrong similarities to each other which they do nothave with other items in the space. Bearing inmind the goal of the work which is to accomplishthe placement of features automatically, thediscovery of these groups is not a trivialprocedure. Some of the problems that must becontended with are: how many of these groupsexist? how do we measure (dis)similarity betweendata items? when do we stop trying to divide thegroups that we already have?

In order to identify districts within an arbitrarydata space we can use techniques from the field ofcluster analysis, for which a number of algorithms

have already been developed. Such algorithmstake collections of data and analyse themaccording to the strengths of similarity of theirdifferent attributes. The output a clusteringalgorithm is a set of discrete clusters of data itemswhich we directly map onto the idea of differentdistricts in a visualisation of this data. Oncedistricts have been identified, they need to berepresented in the visualisation. For example, onemight use the attributes of colour and shape togive each district a distinct character (see belowfor examples). Districts also provide the basis forcreating the remaining legibility features.

It should be remembered however that theprime purpose of the representation of districts inthe information visualisation systems is to providea segmentation of the space which can be used asa reference for navigation. Therefore, while it isdesirable for the districts to have a semanticrelevance the strict accuracy of the classificationmethod need not be an overriding concern. Weaccept that this is a major research issue in itselfand that there are a number of issues which couldbe considered, such as fuzzy boundaries betweenclusters and hierarchical classifications, but ourpriority has been to establish a clustering toolwhich can be applied to a large variety of spaceswith little or no modification. Scope exists toextend and explore the use of other aspects ofclassification methods as further research.

In choosing an algorithm for the initial LEADSimplementation, we set the criteria that it shouldbe simple, due to time constraints on the project,relatively computationally inexpensive andreasonably effective for a wide range of clusters.We have initially adopted Zahn’s MinimalSpanning Tree algorithm (MST) [18]. Thisalgorithm first constructs a minimal spanning treeof the data such that the edge values are taken tobe the euclidian distance, or other similaritymeasure, between the items. Clusters are thenproduced by identifying and eliminatinginconsistent edges, which are defined as edges of

7

the spanning tree whose values are significantlygreater than the nearby edge values. For examplean edge might be eliminated if it its length morethan twice the mean length of other edges withina certain number of steps of if it is more thanntimes the standard deviation of those edgeslongerthan the mean. Useful values forn are usuallybetween 2.0 and 3.0.

The basic algorithm as described by Zahn isable to detect clusters of varying shapes and sizesso long as they are relatively distinct (some of thespecific cases where it will not perform accuratelyare described in Zahn’s original paper). A specificadvantage for LEADS is that the use of minimalspanning trees lends itself to cluster analysis of socalled network structured spaces where one canidentify explicit links between data objects (e.g.hyper-media structures or generalised objectmodels). The drawback to this is that in spaceswhere no such graph structure exists it isnecessary to assume a fully-connected graphstructure, where each object is conceptuallyconnected to all others. This means that the MSTalgorithm must make use of a complete matrix ofdistance values between the objects, making thisstep of the algorithm quite computationallyexpensive. Despite this, and especially in the caseof network spaces, when compared to otherclustering algorithms this method is relativelyinexpensive computationally. Fast algorithms forproducing minimum spanning trees are wellknown and the identification of inconsistent edgesis relatively simple.

A potential problem with the use of aclustering algorithm to discover districts withinthe data is the coherence of the clusters in thespace. If the clustering is based on a similaritymeasure which uses alln dimensions of the data,wheren may be much larger than three, then whenthe data is mapped onto the three dimensions ofthe space then it cannot be guaranteed that theobjects within each cluster will form a coherentgroup. In practice here we are relying on the

efficacy of the underlying visualisation system toprovide a layout for the space which provides ahigh enough degree of proximity between similarobjects. The systems to which we have appliedLEADS so far have shown no obvious signs ofcluster fragmentation. We have also consideredother methods of dealing with this potentialproblem, such as basing the clustering on aweighted combination of the similarity based onthe data and euclidian distance in the space, orusing other clustering algorithms based on morehighly spatial criteria.

4.2. Edges

The next feature we need to cover is the edge.These are (usually) imposing features whichsection off one area from another. It seemssensible therefore to define edges as existingbetween adjoining districts. The main problems tobe solved with edges are: between which districtsshould they be placed? what shape should theybe? how big should they be and how should theybe positioned and oriented?

We will consider three possible methods fordefining edges. The first is a quick and dirtymethod which will not be effective for all shapesof clusters but which might work tolerably wellfor those which are generally spherical or cuboidin shape. This simple approach is to find thenearest neighbour data items between the clustersusing the same similarity measure as wasemployed in the initial clustering process. Theedge can them be placed between the clustersalong the line connecting these two items, with anappropriate orientation. Provided that the objectsdo not excessively cut into the clusters or arepositioned far from their logical joining pointthese methods should be effective in providing areference point and defining the borders ofclusters.

The second method involves finding the hull ofeach district and creating an edge just beyond this.

8

This has a number of drawbacks. Mostimportantly finding the hull of the cluster ofobjects in 3D space is not a trivial task and wouldwork out to be computationally expensive.Considering that the system will already becarrying out a large number of computations inorder to carry out the clustering to producedistricts the trade-off between complexity andaccurate positioning of the edge objects becomesquite important, especially if the situation ofhaving to completely redo the whole legibilityprocess after a large number of changes to thespace looks likely to occur often.

The third method is a refinement of this whichagain involves finding the hulls of the adjoiningdistricts. Once this is complete the edge could bedefined by interpolation between the points alongadjoining edges of districts, as shown in figure 1.Sophisticated interpolation methods mightproduce a smoother edge topology at the expenseof some time, as shown in Figure 1. One possibleproblem with this would be deciding which pairsof data items to use for the interpolation process.Items would need to be in the same general area ofthe space but problems might occur in making thechoice, for example, where the items are densely

Cluster 2

Cluster1

Figure 1: Formation of edge between districts byinterpolation between pairs of adjacent data items

packed. It would also be necessary to considerwhether objects may belong to more than one pairwhere the interpolation process is concerned.

Considering the problems that would comewith attempting to find the hulls of all the clustersfor the initial implementation of the LEADSsystem we decided to use the first methoddescribed here, positioning the edges betweencluster nearest neighbours. This highlighted theproblem of deciding the orientation and size of theplanes used to represent the edges. In orientatingedges the method used is to simply align theobject’s shortest axis with the line joining thespatially nearest neighbours. This provides areasonable orientation in the majority of cases.The size of the edge is dictated by the size of theclusters which it separates. The philosophy used isto make the edge some proportion of the size ofthe smaller cluster in the direction of the twomajor axes of the edge. The proportion currentlyused is 80%, but this may be defined by a systemresource. This means that the edge can neveroverwhelm the smallest cluster of the pair and thatits size is dependent on its orientation to someextent.

4.3. Landmarks

Landmarks need to be stable points within thedata space that can be used as a common referencepoint for navigation. We will illustrate threepossible methods which we have devised fordefining landmark positions.

The methods described here are based on thedefinitions of the districts and will therefore relyon these being quite stable so that the landmarksare not constantly moving as the databasechanges. This should not be a problem in therelatively slow moving type of space for whichthe system has been designed, although it must beexpected that the landmarks will be subject tosome drift over time as the database profileevolves.

9

Method 1: Cluster centroids (see figure 2).Most partitional clustering algorithms will definethe centroid of a group while it is being formed.This centroid is a virtual data element that bestdescribes the cluster as a whole. Our first methodis to simply define landmarks to be the clustercentroids as defined within the clusteringalgorithm. If the algorithm used does not definecentroids it should be possible to simply apply acentroid generation method similar to that whichwould be used with a clustering algorithm to thefinished clusters. A slight variation of this methodmight be to place the landmark at a cluster'sgeometric centre. This method will thereforedefine a single landmark for each of the clusterswhich could act as a beacon for navigation to theheart of the district

The main question to consider is if this is themost suitable place for a landmark. The centre ofthe cluster may be densely populated with dataitems but on the other hand it could just as easilybe rather sparse. In large clusters these landmarksmay easily become lost in the vast group. Wesuspect that this method of placing landmarksmight not take account of the concentration ofclusters and data and so will not necessarilyposition landmarks in areas where externalreferences were most useful. We also feel that thedistricts may already be sufficiently well defined

Figure 2: Cluster centroids as landmarks

Cluster Centroid

Landmark

by their own character and edges so that referencepoints at their centres might not be essential.

Method 2: Cluster intersections (see figure 3).This second method places landmarks wherevermore than two districts intersect. This is a simplemethod that requires identifying which districtsare adjacent and finding a midpoint between theclosest members of the districts. This methodresults in fewer landmarks than simply using thecentroids but those landmarks which are createdappear in the areas of the space where a number ofclusters meet. This implies that these are areaswhere there are a number of clusters bordering oneach other. It is in this sort of area where a stablereference point might help in navigation and so isa better choice for the placement of landmarks.

Method 3: Cluster centroid triangulation. Thefinal method again uses the centroids of thedistricts but will place landmarks at the centre ofthe triangles formed by the centroids of any groupof three adjacent clusters. This is illustrated infigure 4 and will produce approximately the samenumber of landmarks as the cluster intersectionmethod. The position of the landmarks, althoughseemingly quite similar to that produced byMethod 2 is significantly different in that thelandmark is placed in a central position betweenthe points which most typify the districts ratherthan relying solely on the geometric point ofintersection. We suspect that this will give a more

Cluster Centroid

Landmark

Figure 3: Cluster intersections as landmarks

10

even balance to the placement of the landmarksthan simply positioning them at the intersectionpoints and will make the landmark positions moreresponsive to the shapes of the clusters as well astheir positions. This is the method that we haveimplemented in the LEADS system.

It is anticipated that this final method mightgive rise to some drift in the position of landmarkobjects over time. However, if the space satisfiesthe criterion of relative stability over time thismovement should not be enough to render thelandmarks ineffective as reference points.

4.4. Nodes and paths

Nodes and paths are strongly inter-related andwill therefore be considered together. We proposethat paths should be composed of links betweenindividual elements of the visualisation.Eventually, we intend paths to evolve as afunction of use of the space with informationabout movement of users and database accessbeing recorded for this purpose. Two main issuestherefore need to be considered. The first iswhich, if any, elements will be used as initialnodes and paths, before any usage information hasbeen gathered. The second is how the path layoutwill evolve with use. So far, we have implementedalgorithms for the first of these and have

Figure 4: Landmark positions calculated usingcluster centroid triangulation

Cluster Centroid

Landmark

proposals for the second.The choice of initial nodes and paths in an

unused information space is a difficult problem.On one hand it will be useful to have thesefeatures available to aid initial legibility andimageability of the space while on the other handwe must be careful not to colour the usage of thespace too much. Presuming that absolutely nousage information has been recorded, it will benecessary to make educated guesses at dataelements which might possibly become nodesthrough usage. We have implemented thefollowing approach in LEADS:• The nearest neighbour elements between

districts might represent good initial choices asthese are the most similar items across districtboundaries. We shall call such nodes gateways.

• An additional main node for each district mightbe defined as that data element closest to thecentroid of the cluster as this might be seen asbeing the item most typical of the district. Innetwork type spaces there may be further dataavailable to identify such nodes automatically,such as connectivity information (valency) ormeasures of amounts of data stored, both ofwhich imply importance.The current LEADS implementation first

identifies the gateway nodes and forms pathsbetween nearest neighbour pairs in adjoiningdistricts and then within districts all the gatewaynodes are linked together with a spanning tree.

Considering dynamic usage information, wepropose that the positions of node, path andgateway features should evolve in response to theway in which they are accessed. The type ofstatistics that will be used will be concerned withboth the frequency of access to individual itemsand the sequence in which the items are examined.For example:• The ability of a data item to potentially become

a node might depend on the number of querieswhich were made to the item within a certaintime period. For example, within each district

11

this metric might be the topn% of thefrequently most used nodes. In network spaceswe might also consider the number ofconnections which lead to and from the item,itsvalency, as this may also be considered as anindicator of the importance of the object in thespace.

• The definition of a path link between twoelements might depend on the number of timesthose elements were accessed in sequence. Forexample, we might define the set of path linksas the topn% of possible links wheren will befound by experimentation.In LEADS, we produce two levels of path

which reflect different levels of use within thespace. Within the set of path edges some of thosemost used will be further highlighted to show themost popular routes between gateways. Specialcare must be taken however to evolve slowly fromthe set of initial paths and nodes. We do not wishto eradicate any legibility features that we havedefined too quickly, but it is also desirable thatpaths and nodes do not remain in the space whenthey do not accurately reflect its usage. It will alsobe important to ensure the connectivity of paths toall gateway nodes.

Other features

Text is an essential feature of manyvisualisations. In information spaces, text may beused as an identification sign of an area orelement, as a sign giving directions to other areasor as an annotation. In his book Wayfinding inArchitecture, Passini talks at length about thenature, placement and content of signs [13] andwe have adopted his philosophy where possible.This book contains an in depth examination of thewayfinding process and concludes that signsshould be placed at points in the space wherewayfinding decisions are necessary and also atregular intervals between these point where thedistance is large as a reassurance to users. In an

urban environment a decision point wouldgenerally correspond to an intersection on a pathso we are instantly provided with a possibility forsign positions in the information space.

In three dimensional informationvisualisations it is often difficult to maintainorientation, and even more difficult to apply thisorientation to the data items in the space torecognise how their attributes are mapped onto themajor axes. A useful addition to the space mightbe an optional representation of the axes whichwill display their orientation with respect to theuser and a textual annotation indicating whichattributes are mapped onto the axes in informationspaces where this is relevant.

Finally, a simple mechanism for realigning theviewpoint with the major axes can easily be addedto the VR visualisation system to allow users toeasily correct unintentional rotations and to regaina familiar orientation.

In summary then, this section has proposed anumber of techniques for generating or enhancinglegibility features in information visualisationsand has described how these have beenimplemented in the LEADS system. Thefollowing section now describes three exampleapplications of LEADS to different pre-existinginformation visualisations.

5. Four example applications

We have applied LEADS to four differentvisualisation systems, three of which were locallyavailable at the start of this research and a fourthwhich was developed externally. These are:• The FDP-Grapher tool for visualising three

dimensional network structures [19];• The VR-VIBE system for interactive

visualisation and searching of documentdatabases and the World Wide Web [20];

• The Q-PIT 3-D scatter-graph visualiser [20];and

12

• The BEAD document space visualisationsystem [10][11];Between them, these four examples cover a

range of different visualisation layouts which arebroadly representative of the different approachesdescribed in the literature. They also includeexamples of fully 3-D visualisations where thedimension of height is used in the same way asdepth and width and also a landscape stylevisualisation where information is arrangedmostly in a plane and the dimension of heightenables the user to fly over the scene and to zoomin and out.

We will now focus on each of theseapplications in turn, describing the underlyingvisualisation system and demonstrating theeffects of applying LEADS to its output. We willgive a particularly detailed description of theFDP-Grapher example, showing images of thevarious different legibility features in isolation, asthis was the example used for experimental work.We will provide somewhat shorter descriptions ofthe other examples, limiting images to moregeneral before and after shots.

5.1. FDP-Grapher

FDP-Grapher is a 3-D tool for visualisingarbitrary network structures. The underlyingvisualisation approach is based on the ForceDirected Placement (FDP) technique where thenodes of the network are treated as masses and thelinks as springs [21]. Initially the nodes are placedrandomly and the system then passes throughrepeated cycles of repositioning the nodes, basedon the tension in the links, until a relatively stableformation is found. In addition, each link in thenetwork can be given a weight which will alter theway in which the tension value is calculated. Theresulting visualisation shows the network drawnin 3-D space such that densely inter-linked groupsof nodes are positioned closely together.

FDP-Grapher might have many applications.

One of our current applications is to visualiseregions of the World Wide Web; in other words,to produce the kind of overview diagramsproposed by Nielsen (see section 2). Users of thisapplication are able to:• see an overview of up to a hundred or so linked

Web nodes defined by an initial position(specified by a WWW URL) and a linkadjacency distance;

• navigate the resulting visualisation with sixdegrees of freedom;

• select nodes in order to either obtain summaryinformation or to launch the Mosaic browser inorder to inspect their contents; and

• grab nodes and reposition them in order tostretch out the visualisation;The WWW is seen as an ideal underlying

database for this example because it satisfies all ofthe criteria for an appropriate space as defined insection 4. The WWW is highly persistent and isre-used on a very regular basis by a large numberof users. Despite its perceived chaotic nature, theWWW as a whole also satisfies our criterion forstability, as the number of significant changesmade to its contents over short periods of time arevery small with respect to the overall size of thedatabase and the number of retrievals.

Figure 5 shows a number of differentscreenshots from FDP-Grapher being used tovisualise a network of 239 nodes (the examplenetwork actually used for experimentation). Eachscreenshot is taken from the same point of viewbut shows different combinations of legibilityfeatures (LEADS allows the user to interactivelyswitch different features on or off). Figure 5 (a)shows the basic visualisation with no additionallegibility features. Figure 5 (b) shows districtsdistinguished by the colour and shape of theirconstituent nodes. Figure 5 (c) shows theplacement of a number of landmarks within thevisualisation and figure 5 (d) shows how thoselinks representing key pathways are emphasisedthrough changes in colour and thickness. Figure 5

13

(a) (b)

(d)(c)

Figure 5: LEADS applied to FDP-Grapher showing individual legibility features:(a) no features, (b) districts, (c) landmarks, (d) paths

14

(g)

(f)(e)

Figure 5:(cont.) Spaces with individual legibility features: (e) nodes (indicated byincreased size of data items), (f) edges, (g) all features together

15

(e) shows the emphasis of key nodes by increasingtheir size and figure 5 (f) includes a number ofadditional edge objects between the majordistricts. Finally, figure 5 (g) shows thevisualisation with all of the features turned onsimultaneously. When studying these images, it isimportant to bear in mind that they represent adistant perspective view and that the user is ableto fly right into the centre of the visualisation, inwhich case they are surrounded by the nodes andlinks. Thus, what might appear to be a somewhatcluttered image from this distance becomes moreopen from closer in or even inside.

5.2. VR-VIBE

The VR-VIBE visualisation [20] supportsinformation retrieval from electronic documentrepositories and is a three dimensional andinteractive extension of the original 2-D VIBE asreported by Olsenet al. [22]. Unlike traditionaltext retrieval systems which only allow users torun a single keyword search at a time, VR-VIBEallows users to explore the effects of comparingand dynamically manipulating multiplesimultaneous keyword searches. In essence, anumber of keyword searches are defined, each ofwhich consists of one or more text keywords.These are then positioned in a virtual space toform a spatial framework of queries. Documenticons are positioned within this frameworkaccording to the strengths of their relativeattractions to each query (i.e. the more strongly anindividual document matches an individual query,the closer it is placed to it).

VR-VIBE users may dynamically interact withthe visualisation in a number of ways: selectingdocuments displays summary details or launchesMosaic to view the document source if stored inthe WWW (links are maintained from the localdocument repository into the Web); raising arelevance filter removes all documents whoseoverall score falls blow a threshold value from the

display; grabbing and dropping queriesdynamically re-arranges the space into a newconfiguration; switching queries on and off alsochanges the space and, finally, any number of newqueries may be defined and positioned on the fly.

Figure 6 shows example screenshots of VR-VIBE before and after the application of LEADS.The first image shows a visualisation ofapproximately 1500 document referencesstretched out within a framework of five queries,currently positioned at the corners of an invertedpyramid. The second image shows the samevisualisation from the same viewpoint but withthe addition of legibility features.

5.3. Q-PIT

The Q-PIT visualiser draws three dimensionalscatter graphs of tabular data. Q-PIT takes datawhich consists of records, each of which as a fixednumber of typed fields, and maps these onto dif-ferent display attributes of a three dimensionalvisualisation. These include the three spatial di-mensions as well as representational attributessuch as colour, shape, size and spin-speed. Q-PITusers are able to specify different mappings be-tween the underlying data attributes and the visu-alisation attributes via a configuration file eachtime they launch the visualisation. Once the visu-alisation has been created, they are then able to in-spect the contents of objects and also modifythem, in which case they may see an animatedmovement of any changed objects to new loca-tions in space. As with 2-D scatter graphs, theremight be many potential uses of a Q-PIT style vis-ualisation involving comparison and correlationof different combinations of attributes belongingto a set of data objects. Figure 7 shows before andafter images of applying LEADS to Q-PIT (in thiscase a small database of some 130 personnelrecords).

An interesting issue raised by this example isthe likelihood of LEADS overloading visual

16

Figure 6: VR-VIBE before and after the application of LEADS

17

Figure 7: Q-PIT before and after the application of LEADS

18

display attributes currently used by Q-PITresulting in possible confusion (e.g. LEADSusing colour to distinguish districts when Q-PIThas already assigned it some other intrinsicmeaning). Unlike FDP-Grapher and VR-VIBE,Q-PIT does not fix the meanings of its displayattributes until run-time. LEADS therefore cannotguarantee that the attributes that it uses toemphasise districts, nodes and paths, or theobjects that it introduces to represent edges andlandmarks will not be confused withrepresentations of the underlying data. Thissuggests that more careful consideration must begiven when applying LEADS type systems tovisualisations that make extensive and dynamicuse of different display attributes. It also suggeststhat visualisations might usefully make suchmappings explicitly visible (as Q-PIT does in itsconfiguration file).

5.4 BEAD

BEAD is a visualisation system developed byMatthew Chalmers at Rank Xerox EuroPARCand continuing at the Union Bank of Switzerland.It was initially designed for the visualisation oflarge document stores but has recently beenapplied to other data such as groups of time seriesrepresenting stock performance. The algorithmsdeveloped by Chalmers for the placement of dataitems in the space are based on annealingmethods. The goal of the algorithm is thereduction of the total energy of the system byincremental movements of objects, where theenergy value is provided by attractive/repulsiveforces between items. The level of the force iscalculated using a similarity measure between theitems, for example word co-occurrence indocument spaces, and their distance from eachother, with similar objects being drawn towardseach other and dissimilar ones forced apart.

A major feature of BEAD is that it uses themetaphor of a landscape to represent the space. To

achieve this the forces in the placement algorithmare skewed so that the space converges towardsthe x-z plane. When the energy of the systemreaches a sufficiently low level the positions ofthe objects are used as the data points for a DeLaunay triangulation which generates thepolygons of the landscape. The result of this is asingle, well defined ‘island’ of data whoseshoreline can be considered as an edge, in themanner of Lynch, and where similar objectsshould be drawn together into district like groups.

The main difference in the application ofLEADS to BEAD spaces is that in the this case thepresence of the landscape object which isgenerated by the system provides scope for therepresentation of legibility features which is notpresent in the other, more generally threedimensional, visualisations considered so far. Wetook advantage of this to improve the cohesionand visual definition of the districts beingdisplayed in the space. The previous exampleshave shown districts which were distinguishedfrom each other visually by changing the shapeand, in particular, the colours of the objects theycontain. For the BEAD system the districts wereshown by colouring the polygons of the landscapeitself rather than the objects. As figure 8 showsthis results in extremely good definition of thearea of the districts. The polygons which fallbetween districts (as is inevitable when using a DeLaunay triangulation) are given a neutral colour.In this way the colouring system also implicitlydefines the edges which emphasise the districtborders.

6. Experimental Evaluation

In this section we describe an initialexperiment which was carried out to evaluate theeffectiveness of the LEADS implementation ofour proposed legibility techniques. The aim of theexperiment was to test the way in which its

19

Figure 8: BEAD before and after the application of LEADS

20

subjects learned the layout of an informationspace which they entered and used repeatedly. Tofacilitate this the subjects were asked to completea search task a number of times in the same space.Separate experiments were carried out in spaceswith and without the LEADS legibilityenhancements. The experiment only consideredthe addition of all five of the legibility featurestogether and not the effects of individual featuresor smaller combinations of features.

The search task was to find five named objectsin a graphical space of approximately 240 nodes.The graph used was randomly generated but notso random that objects of similar numeric valueare spread throughout the space. This meant thatthe search task was not entirely one of finding aneedle in a haystack. The subjects were requiredto complete the search task three times in thespace of two days. The data items for the searchwere chosen to be spread quite evenly through thespace.

From this we can identify that the independentvariable in the experiment was the level oflegibility information added to the basic space.The dependent variable was the time taken tocomplete the search task, which was dependent onhow well the positions of the objects werelearned.

Due to the nature of the environment withinwhich the work was carried out the best source ofsubjects for the experiment was the studentswithin the Department of Computer Science at theUniversity of Nottingham. Apart from the simpleavailability of people this did have anotheradvantage in that it limited the amount ofvariation between the subjects quite significantly.They all had a similar standard of education andof experience with computers and databases. Thesubjects were divided into two groups andallocated randomly to each of the two tasks.

Two types of result were gathered from theexperiments: statistical, from the times taken tofind each object; and anecdotal from observation

of the subjects and through questionnairescompleted after the final attempt at the task.Because the time available for initialexperimentation was limited and hence the size ofthe sample was small (six people in total) thestatistical results cannot be taken as statisticallysignificant. However, in conjunction with theother results we believe that they can provide auseful initial evaluation to guide furtherdevelopment and more detailed evaluation.

6.1. Measured results

These results are based on the times taken bysubjects to find each of the five target objects andhence the time taken to complete the task as awhole, if achieved. Non completion of the taskwas due to an absolute limit of forty-five minuteson each attempt which was seen as a reasonabletime to avoid undue boredom and frustration inthe subjects.

Table 1 shows the absolute times taken forsubjects to find each object (in the correct order)during the trials. The numbers in the left handcolumn indicate the number of the subject and oftheir attempt at the task. Subjects S1, S2 and S3were using the space without any legibilityenhancement. The remaining subjects used thespace with all LEADS features added. Theremaining columns each represent one of the fivetarget objects. Dark shading of the table cellindicates that the target was not found. The mainobservation from the data is that the rate of taskcompletion on the second and third attemptsamongst users of the space with enhancement washigher than for those without. Users of the spacewith enhancements seemed to complete the taskalmost trivially on the third attempt while those inthe raw space still had problems. This would seemto imply that greater learning of object positionsdid take place with the aid of legibility features.

21

Trial T1 T2 T3 T4 T5

S1.1

S1.2 9:00 12:30 25:00 39:00

S1.3 19:00 22:00 41:00 42:30

S2.1 32:00

S2.2 34:00

S2.3 9:00 18:00

S3.1 8:0

S3.2 1:30 4:00 15:00 21:00 32:00

S3.3 10:00 17:00 24:00 25:30 29:00

S4.1 16:00 37:00 44:00

S4.2 2:00 4:30 12:00 15:00 31:00

S4.3 0:30 5:00 9:00 34:00

S5.1 2:00 7:00 17:00 21:00 34:00

S5.2 0:30 2:00 2:30 10:00 14:00

S5.3 0:30 6:30 8:30 12:00 18:00

S6.1 13:00 33:00 34:30 45:00

S6.2 4:00 6:30 7:00 10:00 11:00

S6.3 13:00 14:00 15:00 16:00 17:00

Table 1: Absolute times taken by subjects to find search objects

Another interesting point is that the mean timetaken to find individual objects was consistentlysmaller onall attempts for the users of the spacewith legibility enhancements. These users seemedto gain some immediate advantage from thelegibility features, without the chance for learningto take place. One explanation for this might bethat the existing partition of the space into districtslead to the use of a more structured searchingtechnique even before the space has been learned.

6.2. Results from questionnaires and ob-servation

One of the first things that becomes clear fromviewing the subjects performance was theimportance of the way in which the volunteersmoved through the space. There seemed to be agreat variety in this despite the small sample usedand the way they moved seemed to have aninfluence on the search methods used. The mosteffective search technique seemed to be one of

22

flying through strips of the space and gatheringinformation on the numbers of the objectscontained there. In contrast to this one subjectadopted a technique of moving directly toindividual objects, looking in the immediatevicinity and then pulling right back from the spacebefore moving to another object in a differentsection of the space. This seemed to waste a greatdeal of time and did not allow him to observe asufficient number of objects to be effective for thetask, and this was reflected in his performance.

In general the subjects using the space withlegibility enhancements seemed to remember thepositions of the objects more effectively. Theycommented that the main aid to remembering theposition of the objects was the colour (and to alesser extent shape) of the districts. Althoughthese were the main feature used as a memory aidfrom the observation it was clear that otherfeatures were also being used. One example wasan object which was near to the largest Edge,which appeared as a green plane in thevisualisation. When searching for this target thesubjects often checked the objects along the lineof this plane. Subjects using the space withoutlegibility enhancements would often show signsof remembering certain attributes of the area ofspace the object was in but not the definiteposition.

The questionnaire contained 15 closed and fiveopen questions which fell into five areas designedto gather information on different aspects of theuser’s experience of the space. These categorieswere: orientation, (learning) position of targetobjects, movement, system performance and useof other features (e.g. text labels.) Table 2 lists theclosed questions and the answers given.

The responses to the section regardingorientation in the space are somewhat ambiguous.On the general question of whether the subject feltdisoriented in the space the surprising answer isthat the group without legibility enhancementsfelt less disoriented. This may indicate that the

extra information added to the space couldpossibly add to initial overload of information.However, on the related question of whether theusers felt most disorientation within the group ofobjects or outside it (i.e. gaining an overview) thesubjects from the group with enhancementsseemed more comfortable with the latter, a viewof the whole space. We might surmise then thatthe divisions presented by the clustering intodistricts are of most use on the global scale of thespace and that more effective local cues fororientation are necessary.

The greatest agreement among the volunteerscame in the section on the use of text. All subjectshad text visible almost all of the time. Thesubjects also had strong agreement that the abilityto align the labels towards the current viewpointposition was very useful. This seemed to bereflected in the number of times that the featurewas used during the experiments.

The most important section of thequestionnaire is that on the way in which thesubjects learned the positions of target objects. Inonly one of these (“During the third visit I had noimpression of the overall structure of the space.”)did the subjects show significant agreement in theanswers given. The responses of both groupsindicated that they had in general gained a goodimpression of the overall structure of the space.

However, the questions where the answers diddiffer are more interesting. The first of these wasasking if the subjects knew where the searchtargets were when they attempted the trial for thethird and final time. The subjects using theenhanced space were much more positive thatthey had learned the absolute locations of theobjects than those without enhancements. Theother two questions were testing whether thesubjects knew the positions of the targets relativeto each other. The subjects with legibility featureswere again more positive that they knew therelative direction of their next target object andthat they had learned routes between the objects.

23

Statement Group 1 Group 2

I did not feel disoriented in the virtual space 2 1 4 4 4 2

I made use of the axis object to check myorientation (1=Never, 5=Always)

3 2 1 3 3 2

I considered space to have a fixed top andbottom

5 4 2 2 5 3

I felt less disorientated outside the group ofobjects then when in the centre of the group

2 3 5 1 2 2

During the third visit to the space I knew wherethe search objects were

4 3 5 3 2 2

Navigation in the world was straightforward 2 2 4 2 3 1

During the third visit I had no impression of theoverall structure of the space

4 2 4 4 5 4

I easily learned routes between different searchobjects

4 4 5 2 3 2

During the third visit I did not know in whichdirection to move to reach the next searchobject

4 2 2 5 4 5

I preferred moving freely through the space tofollowing links

1 1 5 3 1 1

The response time of the system was adequatefor my participation

4 4 2 3 4 2

I found it useful to be able to switch betweenthe two modes of movement

2 1 2 1 1 5

I often wished to move more quickly throughthe space

3 2 2 2 2 4

I rarely aligned the text towards my position inthe space

5 4 5 4 5 5

I often had the text visible (5) 2 1 1 1 1

Were you using the space with or withoutdifferent coloured areas? (1=With, 2=Without)

2 2 2 1 1 1

Table 2: Results from closed questions. Group 1 contained the subjectsusing the space without legibility enhancement, Group2 the subjects usingthe space with all features added. Answer 1 = strong agreement, 5 = strong

disagreement, unless stated.

24

The results from this final set of questionsstrongly indicate that the volunteers who wereusing the space with legibility enhancements hadgreater success in learning the positions of theobjects and relating this to their current position inthe space. This meant that they knew both thelocation of the objects and how to reach them.While the subjects using the space had an equallygood impression of the structure of the data theywere not so confident of the positions of theobjects. This therefore provides some support forthe hypothesis that the LEADS legibilityenhancements are useful in helping users inlearning the layout of abstract data spaces.

The open questions highlighted the problemsof using text in three dimensional spaces, such asocclusion by other objects and alignment. Thisseemed to be the major obstacle to efficient use ofthe system.

Subjects using the space with legibilityenhancements also emphasised again theeffectiveness of the districts in helping them toquickly return to objects they had discoveredpreviously.

In summary then, the experiment seems toprovide some initial evidence for the effectivenessof the legibility enhancements for allowing theusers to learn the space. While the emphasis wason the use of districts the other features may haveplayed a supporting and reinforcing role toincrease the overall effect.

7. Conclusions, furtherreflections and future work

This paper has proposed a number of generaltechniques for improving the legibility of threedimensional information visualisations so thattheir users might more easily learn to navigatethem. This work has adapted well establishedtechniques from the discipline of urban planningand, we argue, points in a more general way to a

potentially useful relationship between thesedifferent fields. The primary goal of our work hasbeen to develop a set of algorithms forautomatically creating or enhancing legibilityfeatures within information visualisations. Theseinclude:

• the use of clustering algorithms to createdistricts;

• the creation of edge objects separating dis-tricts;

• the placement of landmark objects at a cen-tral point between districts; and

• emphasising nearest neighbour node objectswithin districts and creating paths betweenthem.

We have implemented these techniques in aprototype system called LEADS which isintended to provide an additional legibility layersitting on top of current informationvisualisations. So far, we have applied LEADS tofour existing and contrasting informationvisualisation tools. The results of initialevaluation work suggest that this approach ispromising and that it warrants furtherinvestigation.

We now conclude this paper by offering someadditional observations which have arisen fromthe process of designing and implementing theLEADS system. We will also propose somebroader directions for future research. We beginwith some specific technical issues.

First, the idea of introducing legibility featuressuch as paths seems to sit uncomfortably with sixdegrees of freedom navigation. Paths in a city areactually travelled along and the environment istherefore experienced from the perspective of thepath. We suspect that users of our legible virtualenvironments should also directly experiencepaths as part of navigation. Thus, we haverecently incorporated a simple interface techniqueto support navigation via paths, so that selecting apath causes the user to traverse it to its destination.

Second, automatically determining an

25

appropriate scale and appearance for legibilityfeatures has not always been easy. In particular,features such as landmarks and edges must bevisible without being intrusive. Creating usefuledges has proved to be a particularly difficult taskas an edge should ideally follow the contours of adistrict. In a 3-D space, determining a sensiblesize for edges has been difficult. At one extremean edge might be a hull completely surrounding adistrict. At the other it might be a thin flat surfacedividing two districts. The former is likely to bevisually intrusive; the latter may provide aninsufficient sense of separation between districts.

Third, other features and tools are clearlyneeded to help people navigate. For example, theuse of textual information in the form of signpostsis an important part of navigating conventionalurban environments. In the field of city planningsignposts are often considered to be something ofan admission of failure [16], that the planner wasunable to produce a sufficiently legibleenvironment, but we concur with the view ofPassini that they are in fact a useful, eveninvaluable tool during wayfinding tasks. For thisreason we are endeavouring to include signpostsin our environments where possible. Currentlythese appear attached to the paths near to the nodeobjects and identify the item to which the pathleads. However, this gives rise to the problem ofhow to name districts and landmarks. Morespecifically, given that districts and landmarks areautomatically created by LEADS, we are left withthe problem of automatic name generation oralternatively the automatic creation of symbolicidentifiers on signposts.

Fourth, we note again the problem ofoverloading visual attributes as discussed insection five. Ideally, LEADS ought to be able toquery a visualisation application to find out whatmappings or representations it already uses, inorder to chose non-conflicting ones for legibilityinformation. In turn, this suggests thatvisualisations ought to make their mappings

available to other applications in some explicitform.

Fifth, we need to be careful that by addingadditional objects to an information visualisation,we do not increase the rendering overhead therebydegrading system performance. However, wesuspect that some LEADS extensions might leadto improved system performance. For example,LEADS might enable the development of adistancing technique for data spaces whereby allof the individual data items in a district would bereplaced by a single object when viewed from adistance.

We also need to instigate a more detailedprogram of experimental work in order to morerigorously test our hypotheses and also to teaseapart the effects on navigation of differentlegibility features and of different techniques forenhancing or creating them.

Adopting a more general perspective, thispaper provides one example of how the field ofurban planning can inform the design of threedimensional information visualisations. However,Lynch’s work on legibility, although very wellknown, is not the only research in this field thatmight be adapted to this purpose. In particular,work on the social logic of space [23] hasconsidered how urban form relates to bothnavigation and social action. Thus, as opposed tofocusing on the relationship between a single userand an information visualisation, we mightinstead consider how to support shared andcooperative access to such visualisations asproposed in recent work on Collaborative VirtualEnvironments and Populated InformationTerrains [20]. Some work into this topic hasrecently been carried out by John Bowers as partof the European COMIC project resulting in aprogram called the Virtual City Builder whichconstructs virtual cities based on underlyingprinciples proposed by Hillier and Hanson. One ofthe main aims of our future research is to explorethe integration of this work with our own

26

Legibility based techniques in order to gain amore complete view of how urban planningtechniques can support the design of futurevisualisation systems. An initial discussion of therelationship between these approaches can befound in [24].

8. References

[1] Elm, W. C. and Woods, D. D. (1985)Getting Lost: A Case Study in InterfaceDesign.Proceedings of the Human FactorsSociety. pp. 927-931.

[2] Nielsen, J. (1990) The Art of NavigatingHypertext.Communications of the ACM, 33,3. 297-310.

[3] Nielsen, J. (1990) Hypertext andHypermedia, Academic Press, San Diego,CA. ISBN 0-12-518410-7.

[4] Nielsen, J. (1995) Multimedia andHypertext: The Internet and Beyond.Academic Press, SanDiego, CA. ISBN 0-12-518408-5.

[5] Mackinlay, J. D., Robertson, G.G. and Card,S. K. (1991) Perspective Wall: Detail andContext Smoothly Integrated, InProc.CHI’91, Human Factors in ComputingSystems. ACM SIGCHI, May 1991, pp 173-179.

[6] Card, S. K., Robertson, G. G. andMackinlay, J. D. (1991) The InformationVisualiser, an Information Workspace. InProc. CHI’91, Human Factors in ComputingSystems. ACM SIGCHI. May, 1991. pp 181-188.

[7] Erikson, T. (1993) Artificial Realities asData Visualisation Environments: Problemsand Prospects. In Alan Wexelblat (ed.)Virtual Reality: Applications andExplorations. Academic Press Professional,Cambridge, MA. ISBN 0-12-745045-9. pp.3-22.

[8] Fairchild, K. M. (1993) InformationManagement Using Virtual Reality BasedVisualisations,. In Alan Wexelblat (ed.),Virtual Reality: Applications and

Explorations. Academic Press Professional,Cambridge, MA, ISBN 0-12-745045-9, pp.45-74.

[9] Darken, R. P. and Sibert, J. L. (1993) AToolset for Navigation in VirtualEnvironments. InProceedings of UIST'93,November 3-5, pp. 157-165.

[10] Chalmers, M. (1993) Using a LandscapeMetaphor to Represent a Corpus ofDocuments. InProc. European Conferenceon Spatial Information Theory, Elba,September. Published as: Frank, A. andCampari, I. (eds.),Spatial InformationTheory. Springer Verlag. pp. 377-90.

[11] Chalmers, M. (1995) Design perspectives invisualising complex information. InProc.IFIP 3rd Visual Databases Conference,Lausanne, Switzerland, Chapman and Hall.

[12] Lynch, K. (1960)The Image of the City.MIT Press, Cambridge, MA.

[13] Passini, R. (1992) Wayfinding inArchitecture. Van Nostrand Reinhold.

[14] Piaget, J. and Inhelder, B. (1969)ThePsychology of the Child. Routledge andKegan Paul. First published in French as LaPsychologie de L’enfant. PressesUniversitaires, Paris. 1966

[15] Downs, R. M. and Stea, D. (eds.) (1973)Image and Environment: CognitiveMapping and Spatial Behaviour. AldinePublishing Company, Chicago.

[16] Canter, D. (1984) Way-finding andSignposting: Pennance or Prosthesis. InEasterby and Zwaga (eds.),Nato Conferenceon Visual Presentation of Information,Information Design.Wiley.

[17] Ingram, R. J. and Benford, S. D., Improvingthe Legibility of Information Visualisations.In Proc. IEEE Viz’95. Atlanta, US.November 1995, IEEE Press.

[18] Zahn, C.T., Graph-theoretical Methods forDetecting and Describing Gestalt Clusters.IEEE Transactions on Computers. 20, 68-86.

[19] Benford, S., Bowers, J., Fahlén, L. E., et al.(1995) Networked Virtual Reality and Co-operative Work.Presence: Teleoperatorsand Virtual Environments, 4, 4, 364-386.

27

[20] Benford, S., Snowdon, D., Greenhalgh, C.,et al. (1995) VR-VIBE: A VirtualEnvironment for Co-operative InformationRetrieval.Computer Graphics Forum. 14, 3.349-360.

[21] Fruchterman, T. M. J. and Reingold, E. M.,(1991) Graph Drawing by Force DirectedPlacement. Software Practice andExperience, 21, 11, 1129-1164.

[22] Olsen, K. A., Korfhage, R. R., Sochats, K.M. et al. (1993) Visualisation of a DocumentCollection: The VIBE System.InformationProcessing and Management, 29, 1, 69-81.

[23] Hillier, B. and Hanson, J. (1984)The SocialLogic of Space. Cambridge University Press,Cambridge.

[24] Ingram, R. J., Benford, S. D. and Bowers, J.M. (1996) Building Virtual Cities: ApplyingUrban Planning Principles to the Design ofVirtual Environments In Proc. ACMConference on Virtual Reality Software andTechnology (VRST’96).Hong Kong, July1996. pp 83-91. ACM Press.