Visualization of Large Ontologies with Landmarks

Zong Lei Jiao

, Qiang Liu

, Yuan-Fang Li

, Kim Marriott

1,2

and Michael Wybrow

1,2

Faculty of Information Technology, Monash University, Clayton, Victoria 3800, Australia

National ICT Australia, Victoria Laboratory, Melbourne, Australia

Keywords:

Ontology, Visualization, Interaction.

Abstract:

Ontologies are building blocks of the Semantic Web and are essential for knowledge representation and data

integration. Interactive visualization is an important tool for the understanding of ontologies, especially large

ones. In this paper, we present a novel hierarchical ontology visualization method for effectively visualizing an

ontology’s global landmarks, local structure and individual class’ details. We have implemented a visualization

system that can handle large ontologies efﬁciently. Preliminary evaluation indicates that the method and the

system can be effectively used to understand large ontologies.

1 INTRODUCTION

Ontologies underpin the Semantic Web. They de-

scribe domain knowledge using classes (abstract con-

cepts), predicates (binary relationships) and individu-

als (concrete instances). A class represents a set of in-

dividuals and a predicate represents a binary relation

between elements of two classes. Ontologies have

been successfully applied in a number of domains to

facilitate the representation, inference and integration

of knowledge and data (Ruttenberg et al., 2009; Smith

et al., 2007; Bizer et al., 2009).

As ontology use has become more prevalent, the

number and size of ontologies has increased rapidly.

Some of the largest ontologies—such as SNOMED

CT (Stearns et al., 2001), the Galen Ontology (Rogers

and Rector, 1996) and the Gene Ontology (Ashburner

et al., 2000)—contain tens and even hundreds of thou-

sands of classes and predicates. This makes under-

standing, maintenance and application of them a ma-

jor challenge. New methods and systems are needed

that support effective interrogation and comprehen-

sion of such large ontologies.

When understanding a large ontology two of the

most fundamental tasks are: (1) ﬁnd a particular class

of interest and understand its relationship with other

classes in the ontology, and (2) understand the overall

structure of the ontology, in particular the taxonomy

of classes and properties, i.e., their inheritance hierar-

chy.

Information visualization techniques have been

applied to help with these tasks, with a large number

of different approaches proposed over the years (Kati-

fori et al., 2007; Lanzenberger et al., 2010). These ap-

proaches include 2D and 3D visualization, and range

from simple indented list and tree-based layout meth-

ods to space ﬁlling and distortion based techniques.

However, it is fair to say that none of the existing tech-

niques or systems are adequate.

In their comprehensive survey (Katifori et al.,

2007), Katifori et. al. point out a number of common

deﬁciencies of current systems. These include:

Clutter — that the visualization produced is messy

and not easy to make sense of, and made worse

with the increase in ontology size,

Structure — the visualization does not effectively

show the overall hierarchical structure of the on-

tology taxonomy,

Scalability — medium and large ontologies are not

handled efﬁciently, and

Inspection — it is difﬁcult to interactively inspect

and interrogate the ontology in different ways, in-

cluding search, highlight and querying.

In this paper, we tackle the problem of visualizing

large OWL (Horrocks et al., 2003) and OWL 2 (Grau

et al., 2008) ontologies and present a novel visual-

ization method and system that address the deﬁcien-

cies identiﬁed above. Our main contributions are two-

fold:

• The proposal of a novel tri-view land-

mark+local+axiom ontology visualization

model that addresses the problems of clutter and

structure.

461

Jiao Z., Liu Q., Li Y., Marriott K. and Wybrow M..

Visualization of Large Ontologies with Landmarks.

DOI: 10.5220/0004289904610470

In Proceedings of the International Conference on Computer Graphics Theory and Applications and International Conference on Information

Visualization Theory and Applications (IVAPP-2013), pages 461-470

ISBN: 978-989-8565-46-4

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

• The development of a scalable and feature-rich

ontology visualization engine that supports this

model. The engine implements efﬁcient, auto-

matic graph layout algorithms and is capable of

visualizing and interrogating large ontologies efﬁ-

ciently, hence addressing the issues of scalability

and inspection.

Our visualization model consists of three separate

views that allow the user to visualize classes and prop-

erties in an OWL ontology in increasing granularity:

landmark view, local view and axiom view. Firstly,

an orthogonal tree-based landmark view of the class

(resp. property) hierarchy is maintained. This land-

mark view provides a quick global perspective for

the entire ontology. Since it is infeasible to show all

classes for large ontologies, this view displays a sim-

pliﬁed hierarchy using a number of inferred and user-

deﬁned landmark or key classes.

Secondly, a local view presents the full inheri-

tance hierarchy for a small number of classes (resp.

properties). The user can browse within this view by

expanding and collapsing branches of hierarchy, and

by changing the particular class (resp. properties) on

which the local view is focused.

Thirdly, an axiom view shows the OWL axioms of

a particular selected class (resp. property) and the as-

sociated anonymous expressions using a nested rect-

angles visualization based on the expression syntax.

We have implemented our ontology visualization

method as a plugin for the Dunnart (Dwyer et al.,

2008) constraint-based diagram editor. A screenshot

of the system visualizing the well-known Galen On-

tology (Rogers and Rector, 1996) can be seen in Fig-

ure 1. Note that Galen is a large ontology that contains

more than 23,000 classes and 63,000 axioms.

The visualization engine includes a number of

novel features:

• Efﬁcient, automatic layered tree-like layout of the

inheritance hierarchies in the landmark view and

local view.

• Synchronized, bidirectional navigation between

the landmark view and local view.

• Automatic selection and dynamic update of the

landmark classes (resp. properties) in the land-

mark view.

• Flexible search and highlighting capabilities.

• Fast loading of large ontologies via a database

backend.

• Natural visualization of anonymous class (prop-

erty) expressions and axioms in the axiom view.

The rest of the paper is organized as follows. Sec-

tion 2 brieﬂy surveys relevant works. Section 3 dis-

cusses our landmark+local+axiom visualization ap-

proach, including a detailed description of the system

functionality and performance. Section ?? concludes

the paper and presents directions for future work.

2 BACKGROUND

2.1 Semantic Web Ontologies

Mainstream ontology languages such as OWL (Web

Ontology Language) (Horrocks et al., 2003) and

OWL 2 (Grau et al., 2008) are deﬁned based on ex-

pressive description logics (Baader and Nutt, 2003).

The most important kind of object in an ontology

are the classes. These are organized in an inheritance

hierarchy called the class taxonomy. A key require-

ment of ontology visualization is to be able to under-

stand this taxonomy. Ontologies may contain multi-

ple inheritance so this is not a true hierarchy but rather

a DAG (directed acyclic graph). While inheritance

is the primary relationship relating classes, an ontol-

ogy language also includes a number of language con-

structs that can be used to construct complex expres-

sions and axioms to describe and relate classes, pred-

icates and individuals using other properties.

For example, the Campylobacter class in the

Galen Ontoolgy (shown in both Figure 1 and Fig-

ure 2) is deﬁned with the following axioms and class

expressions (in the OWL Manchester syntax):

The deﬁnition of class Campylobacter

Class: Campylobacter

SubClassOf:

Bacterium

hasCellMorphology some

(CellMorphology and

(hasAbsoluteState some flagellated))

hasShape some (Shape and

(hasAbsoluteState some curving))

isActedOnSpecificallyBy some

(Gramstaining and

(hasEffectiveness some

(Effectiveness and

(hasAbsoluteState some ineffective))))

In the above ontology snippet an OWL

class, Campylobacter, is being deﬁned. The

SubClassOf axioms deﬁne four super classes for

this class, three of which are anonymous, nested

class expressions that restrict the membership of

this class. Such class expressions include existential

quantiﬁcation and class conjunction.

The semantics of an ontology is based on its in-

terpretation I (Baader and Nutt, 2003). A class C

is interpreted (denoted C

) as a set of individuals. A

predicate p is interpreted (denoted p

) as a set of pairs

IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications

462

Figure 1: Screenshot of our system being used to visualize the Galen Ontology. The landmark view is shown on the left

(Section 3.1), the local view in the center (Section 3.2), and the axiom view for one class (Campylobacter) in the inset

window (Section 3.3). The information pane of the class Campylobacter is shown in top right, below the canvas zoom

control (Section 3.5). Search and highlighting can be performed in the bottom left panel (Section 3.7), and history navigation

is show in bottom right (Section 3.8).

hx, yi that are mapped under p. The interpretation-

based semantics of the axioms and expressions used

in the above example can be found in Table 1 below.

Table 1: Semantics of some OWL axioms and deﬁnitions.

Axiom/Expression deﬁnition

A SubClassOf: B A

⊆ B

A and B

∩ B

p some C {x|∃y.hx, yi ∈ p

∧ y ∈ C

}

The number and size of ontologies have been

growing steadily over the past decade with the in-

creased adoption of the Semantic Web and Linked

Data (Bizer et al., 2009) techniques. In domains

like bioinformatics, a number of large ontologies have

been used as an important tool for the task of captur-

ing domain knowledge and facilitating knowledge ex-

change and data integration (Ruttenberg et al., 2009;

Smith et al., 2007).

Some well-known and widely-used ones that are

available on the BioPortal ontology repository

in-

clude the Galen ontology, the Gene Ontology, the

NCBI organismal classiﬁcation ontology, the NCI

Thesaurus ontology and the SNOMED CT (clinical

terms) ontology. These large ontologies contain tens

http://bioportal.bioontology.org/

(accessed 2012-11-28)

of thousands up to almost a million classes. Thus,

there is a pressing need for visualization systems to

assist in the understanding and maintenance of such

large ontologies.

2.2 Ontology Visualization

Quite a number of visualization techniques have been

employed in different systems to facilitate the navi-

gation and interrogation of OWL and RDFS ontolo-

gies. A number of ontology visualization systems are

catalogued in (Lanzenberger et al., 2010), with a fo-

cus on tools that support ontology mapping and align-

ment. A comprehensive survey (Katifori et al., 2007)

categories these systems according to their visualiza-

tion approaches. The survey also critically analyses

the visualization systems, and identiﬁes a number of

common ﬂaws. As discussed in Section 1, these short-

comings can be grouped into the following categories:

clutter, structure, scalability and inspection.

In the rest of this subsection, we brieﬂy introduce

the main visualization approaches and discuss their

advantages and shortcomings.

Indented List is a basic, textual visualization ap-

proach that arranges classes (resp. properties) in a tree

view. This approach is employed by most ontology

VisualizationofLargeOntologieswithLandmarks

463

(a) Prot

e’s visualization of the Galen Ontology, showing

the axioms of the Campylobacter class.

(b) KC-Viz’s visualization of the Galen Ontology, showing

the axioms of the Campylobacter class.

Figure 2: Visualization of the Galen Ontology produced by Prot

e and KC-Viz, focused on the same OWL class,

Campylobacter, as in Figure 1.

editors, including the popular Prot

e ontology edi-

tor (shown in Figure 2(a)). The main advantages of

this approach includes intuitiveness, familiarity, non-

overlapping of labels and fast browsing.

A disadvantage is that indented lists are not well

suited to handling multiple inheritance. However,

since multiple inheritance is not that common, some

ontology visualization tools, such as Prot

using

representations designed for hierarchical data handle

multiple inheritance by duplicating the concept and

its sub-concepts in the hierarchy under each of its par-

ents. We believe such duplication is confusing to the

user and leads to an unnecessary increase in the size

of the visualization.

Lack of an overview and the poor use of space

have also been identiﬁed as shortcomings of indented

lists. Nevertheless, the indented list is recognized as a

baseline visualization approach that can be combined

with other techniques. Note that in Figure 2(a) only a

small portion of the hierarchy is shown in the indented

list on the left. Hence, indented lists suffer from the

issues of structure.

Node-link Graphs and Trees are another intuitive

approach to visualizing hierarchies where nodes rep-

resent classes and edges represent inheritance rela-

tionships. It naturally captures the usually hierarchi-

cal inheritance relationship among classes in an on-

tology and for moderately sized networks it can, at

least in principle, clearly show the sub-class relation-

ship and the class structure. The main disadvantage

is that it is not a particularly compact representation

and networks with more than a few hundred textu-

ally labelled nodes are not really understandable when

viewed on a standard screen. Also, depending upon

the layout method, there may be efﬁciency issues for

large ontologies. Hence applications, including On-

http://protege.stanford.edu/ (accessed 2012-11-28)

toViz,

that employ such an approach suffer from

clutter and inspection issues (Katifori et al., 2006).

Zoomable (nested) approaches such as Jambalaya

and CropCircles

place child nodes inside their par-

ents and allow nodes to be zoomed in and out. Be-

ing effective in locating speciﬁc nodes, zoomable ap-

proaches have difﬁculty in handling labels and main-

taining a global overview. Such nested visualization

systems therefore face clutter and structure issues.

Also, in the case of multiple inheritance they require

duplication of a concept and its sub-concepts in the

hierarchy under each of its parents.

Space-ﬁlling techniques aim at making full use of

the display area by ﬁlling the space with nodes. A

node’s display area is subdivided among its descen-

dant nodes, with the size of a node allocated in pro-

portion to the node’s size, number of child nodes, etc.

Systems based on this technique, such as Treemap,

are effective in visualizing ontologies with a shallow

inheritance hierarchy (2 to 3 levels). However, they

are not as effective in presenting the structure of the

ontology. Similarly, they suffer from the issue of

structure.

Focus + Context or Distortion techniques allow

the view to be distorted to focus on a particular node,

while shrinking the other nodes. Such approaches,

http://protegewiki.stanford.edu/wiki/OntoViz

(accessed 2012-11-28)

http://thechiselgroup.org/2004/07/06/jambalaya/

(accessed 2012-11-28)

http://www.mindswap.org/2005/cropcircles/

(accessed 2012-11-28)

http://www.cs.umd.edu/hcil/treemap/

(accessed 2012-11-28)

IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications

464

including Hyperbolic Tree Visualization,

facilitate

easy navigation in the hierarchy while maintaining

contextual information (neighbor nodes). However,

hierarchies, an important concept in ontologies, are

not easily shown in such approaches. These systems

face the problem of structure.

Except Prot

e, all of the above approaches fo-

cus on showing the structure of the ontologies, but not

on showing the details of individual classes (proper-

ties). Hence, they all face the problem of inspection.

Prot

e, being an ontology editor, combines the in-

dented list view with a panel that shows the details of

individual nodes in the list. The details of a class are

shown in the bottom right panel in Figure 2(a).

More recently, concept diagrams have been ap-

plied to visualize a number of different relations of a

group of classes and properties (Howse et al., 2011).

Visual notations are deﬁned for various ontological

axioms and expressions such as class disjointness,

equivalence and all values from. However, this pro-

posed visualization method has not been implemented

yet and it is unclear how effective it will be in han-

dling large groups of classes and properties.

Other recent works including (da Silva and Fre-

itas, 2011; Vercruysse et al., 2012) face similar issues

to those outlined above.

Recently, KC-Viz (Motta et al., 2011), a node-

link-based ontology visualization tool has been de-

veloped as part of the NeON project.

KC-Viz uses

a hybrid representation of an indented list for class

listing and a node-link graph for local structure. The

node-link graph is initialized with a number of key

concepts, concepts that are deemed to be important

by some metrics, to provide an initial, quick sum-

mary of the entire ontology. These behave like land-

marks. However, an overview containing just the

landmarks is not preserved during interaction. Dur-

ing exploration the node-link graph is expanded as

non-key concepts are explored. The key concepts

are not visually distinguished from regular, non-key

concepts. Thus the utility of the key concept view

is limited once more concepts are added to the visu-

alization. The structure and clutter issues intrinsic

to the indented list and node-link graph methods are

also present in KC-Viz. KC-Viz’s visualization of the

Campylobacter class in the Galen Ontology can be

found in Figure 2(b).

Observing that only a handful of visualization sys-

tems are capable of handling ontologies with more

than 100,000 nodes, it is concluded (Katifori et al.,

http://jowl.ontologyonline.org/HyperBolicTree.html

(accessed 2012-11-28)

http://www.neon-foundation.org/

(accessed 2012-11-28)

2007) that scalability is an important issue in ontol-

ogy visualization.

From the above discussion it can be clearly seen

that no single visualization approach effectively ad-

dresses the four deﬁciencies we identiﬁed above. This

observation motivated us to propose a novel tri-view

visualization approach that alleviates such problems.

3 ONTOLOGY VISUALIZATION

WITH

LANDMARK+LOCAL+AXIOM

VIEWS

From the last section, it is clear that a key problem

in ontology visualization is how to provide a mean-

ingful overview of the class taxonomy for large on-

tologies that can be computed efﬁciently and which is

smoothly integrated with detailed information about

a particular class. While many previous approaches

have used techniques designed for single inheritance

hierarchies, we believe that it is important to use a

visualization that can naturally handle DAGs arising

because of multiple inheritance. The most common

representation for a DAG in information visualization

applications is a hierarchical or layered tree-like net-

work diagram. However, as we have discussed, while

well suited to showing taxonomies with a few hun-

dred nodes it does not scale to larger taxonomies.

Because of the very large number of classes in

large taxonomies we do not believe it makes sense to

show all of the classes at once. These taxonomies are

so large that even if each class uses only a few pixels,

this will take up most of the screen space. We be-

lieve a better approach is to use representative classes

in the taxonomy to provide an overview. These rep-

resentative classes act as landmarks for navigation,

and the position of individual classes in the taxon-

omy is understood w.r.t. the landmark classes in the

overview. The use of a landmark overview for navi-

gating in large networks has previously been used for

visualizing the structure of large websites (Mukher-

jea and Hara, 1997). In a closely related idea, the pa-

pers (Peroni et al., 2008; Motta et al., 2011) suggest

providing an abstraction of the taxonomy by showing

the sub-taxonomy for selected key concepts. How-

ever, as discussed in the previous section, the design

of their visualization model differs signiﬁcantly from

ours.

In this section we detail our novel tri-view ap-

proach to ontology visualization including the inter-

action model and the interesting implementation is-

sues we encountered when building a system based

VisualizationofLargeOntologieswithLandmarks

465

on the model. The model has three kinds of views:

the landmark view which provides an overview of

the structure of the class (or property) taxonomy by

showing a sub-taxonomy for landmarks which com-

prise automatically identiﬁed key concepts as well as

user-speciﬁed landmarks; a local view that shows all

concepts in some sub-network of the taxonomy. Both

the landmark and local view of the taxonomy use a

layered network representation. The ﬁnal view, the

axiom view details the properties associated with a

class. The user can open multiple local and axiom

views with different focal areas, if they desire.

3.1 Landmark View

The landmark view consists of the sub-hierarchy con-

necting the landmark nodes. The ﬁrst time that an

ontology is viewed with our tool the initial landmarks

must be determined. This is relatively costly. How-

ever as the landmarks are stored between sessions this

needs only be done once for that ontology.

Our starting point for ﬁnding the n key concepts

for the landmark view is the algorithm given in Per-

oni et al. (Peroni et al., 2008). This uses a set of

static measures to calculate a score for each concept

in ontologies. These measures include natural cate-

gories, density, coverage, and popularity. Natural cat-

egories aim to identify concepts that are information-

rich in a psycho-linguistic sense. Density highlights

the information-rich concepts in an ontological sense.

The popularity criterion aims to identify concepts that

are particularly common. Coverage is used to ensure

that no important parts are ignored. We found that

this algorithm was very slow for large ontologies and

modiﬁed it as follows: we no longer check for cover-

age as this was expensive and in practice makes little

difference; rather than performing a Google search to

ﬁnd out how popular a particular concept is, we use

the number of times it appears in the ontology axioms.

However this static measure is only the initial de-

termination: the choice of key concepts evolves dy-

namically as a result of user interaction. The number

of times the concept is selected in either the landmark

view or local view is tallied to give a good indication

of its importance. Furthermore, we allow the user

to directly promote or demote concepts to the land-

mark view. The ﬁnal importance score for each con-

cept is computed by combining the static and dynamic

scores, then sort the scores and obtains the list of key

concepts.

A subontology for the key concepts is then ex-

tracted. Multiple inheritance in the ontology is re-

moved from the subontology by ﬁrst computing the

depth of each concept from the root and then treat-

Figure 3: Screenshot of landmark view (Galen Ontology).

Non-orthogonal edges indicate multiple inheritance.

ing a concept with multiple parents as the child of its

deepest parent. We compute the layout using Walker’s

layered tree layout algorithm with the modiﬁcations

to ensure linearity (Walker, 1990; Buchheim et al.,

2002). Finally, the removed parent-child edges are

reintroduced into the tree once we have computed the

layout, as shown in Figure 3. Organizing nodes based

on their lexicographical ordering has been shown to

improve the usability of graphs. This is the reason we

use a modiﬁcation of Walker’s algorithm rather than

a standard DAG layout algorithm such as Sugiyama

layered layout (Sugiyama et al., 1981) which reorders

nodes within levels.

We also perform one additional simpliﬁcation of

the resulting tree. When there are multiple leaf nodes

(nodes without any children) of a single parent above

some threshold, we collapse these down to a single

larger node in the landmark view. The main reason

for this is to make better use of space and to prevent

particularly broad trees from being very long in the x

or y dimension.

The landmark tree is scaled uniformly to ﬁt within

the landmark pane of the interface, which is can itself

be resized and repositioned by the user. When the

user hovers over nodes in this view, a tooltip will show

the name of the concept beneath the cursor. Clicking

anywhere in this view chooses a new focal area based

on the landmarks in closest proximity to the mouse

cursor. A square highlight is drawn around landmarks

that are part of the current focal area, as shown with

the root node of the tree in Figure 3.

IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications

466

3.2 Local View

The local view shows a subset of the taxonomy,

which we refer to as the focal area. This is a user-

conﬁgurable amount of concepts, but we prefer a de-

fault of 40 nodes to keep the network understandable

and manageable. The local view is centered around a

single focal concept, either selected by the user click-

ing in the local view, a search result, or the clos-

est concept to the point they clicked in the landmark

view. In order to understand the relationship between

the local view and the landmark view, this node, its

parents and children are shown with a distinct color

coding in both views, as shown in the center pane of

Figure 1. In both views, the focal node Thing are

coded in darker blue whereas its child nodes are coded

in lighter blue. If the focal concept is not a landmark

then the closest landmark to it (graph theoretically) is

color coded in the landmark view. In addition, if the

user brushes (hovers their mouse) over any concept

in the local view, a red circle is momentarily drawn

around the closest corresponding concept in the land-

mark view.

We use constrained network layout (Dwyer

et al., 2006) together with orthogonal edge rout-

ing (Wybrow et al., 2010) to perform the layout of

the local view. For the purpose of layout, we consider

the network in the local view to be a tree, discarding

edges causing multiple inheritance. We then traverse

this tree and assign a level for each node based on its

distance from the root node. Next we perform force-

directed layout, constraining nodes on each level to

be aligned with each other, and ordering between lev-

els (causing the tree to ﬂow in a particular direction).

The initial force-directed aspect of the layout causes

leaves from separate branches to move apart from

each other and not be intertwined (see center pane of

Figure 1). Constraints are effectively used to ensure

that the layout is layered and that the relative place-

ment of concepts shared between the landmark view

and local view is the same. We also apply additional

constraints to prevent nodes from overlapping. This

ensures that their labels remain readable.

Edges between nodes are drawn as orthogonal

routes made of horizontal and vertical line segments

with rounded corners. We position the middle seg-

ment of edges in the channel bordered by the largest

nodes in each level. We also adjust the position within

this channel to distinguish between bundles of edges

connecting common parents or children, as can be

seen between the middle levels in Figure 1.

Figure 4: Screenshot of the Axiom view for the

Campylobacter class (Galen Ontology).

3.3 Axiom View

The concept taxonomy, while important, is not the

only relationship between concepts. A separate ax-

iom view window shows the axioms associated with

the focal concept. More exactly this shows the sub-

, super-, equivalent-, and disjoint-class axioms that

directly refer to that concept. Figure 4 shows this

window for the Campylobacter class, with some sub

classes and (anonymous) super classes. The user can

choose to show the formula textually or using nested

rectangles based on the syntax similar to the repre-

sentation used in Xing for XML queries in (Erwig,

2003).

3.4 Viewing Properties

The properties in the ontology are presented in a simi-

lar way to the concepts: a separate property local view

shows the taxonomy while a property axiom view

shows the domain and range of a selected property

and the properties it is associated with. In theory an-

other landmark view could be provided for properties

but in practice we have found that the complete prop-

erty taxonomy can usually be shown in the local view.

3.5 Information Pane

An information pane shows textual information about

the current focal concept, as shown in Figure 5. We

use it to show statistics about the focal concept, and

information that is difﬁcult to display elsewhere in the

visualization, such as the fully qualiﬁed URI of the

concept.

VisualizationofLargeOntologieswithLandmarks

467

Figure 5: Screenshot of Information Pane (Galen Ontol-

ogy).

3.6 Navigation

The effectiveness of any ontology visualization tool

depends crucially on how easy it is to navigate

through the ontology to ﬁnd the detailed information

the user is interested in while still keeping track of

where they are in the ontology.

The local view is the main way of viewing con-

cepts in the ontology. A new area of focus is trig-

gered by the user double-clicking on a particular con-

cept in the landmark view, local view, axiom view,

property view or search results list. This selected con-

cept is placed in the middle of the local view and its

immediate ancestors and descendants are shown with

it along with the closest key concepts from the land-

mark view. Each node is displayed with a box at ei-

ther end. Where the concept has ancestors or descen-

dants, the node will show a “+” or “-” symbol that the

user can click to show or hide its parents or children.

The changes to the layout due to nodes being added or

removed is smoothly animated. This allows the user

to explore the neighborhood surrounding the concepts

initially chosen to be displayed in the local view.

The user can promote a concept in the local view

to the landmark view. When this happens the layout

for the landmark view is recomputed and the layout

changes are smoothly animated. This is useful if the

user wishes to provide their own points of interest, or

wish to augment the inferred key concepts with their

own prior knowledge.

3.7 Search and Highlighting

The tool provides a powerful search mechanism. This

is accessible to the user via the Search pane, as shown

in Figure 6. The user can enter a search term. The

system will return a list of classes, individuals and

properties that match the user’s search term. All these

Figure 6: Screenshot of Search Pane (Galen Ontology).

Figure 7: Screenshot of Landmark View with highlighted

search results (Galen Ontology).

matches will be highlighted in both the landmark view

(see for example Figure 7) and the local view. The

user can hover their mouse cursor over any item in the

search results list to cause only that item to be high-

lighted in the landmark and local view (or the closest

concept to it in the case of the landmark view). The

user may click on any item in the search results list to

make that the focal concept and cause the local view

to update to a new focal area. This dual search design

allows the user to ﬁnd and select any concept within

the entire ontology but can also be used to effectively

search and highlight information already shown in the

local view without causing that view to switch to a

IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications

468

Table 2: Loading times for various ontologies (in msec.) from a database.

Name Nodes First load Ontology data Key concepts Drawing Total

ICNP 3,291 yes 1,489 27,700 641 29,830

no 1,479 266 748 2,493

Galen Ontology 23,142 yes 7,214 636,541 822 644,577

no 7,075 7,459 696 15,230

Gene Ontology 35,228 yes 11,215 1,603,332 2,180 1,616,727

no 10,751 96,919 2,797 110,467

different focal area.

3.8 Navigation History

The system also offers a Navigation History pane.

This pane shows an ordered list of concepts selected

via the landmark view, local view, axiom view, prop-

erty view or search results list when determining a

new focal area for the local view. This allows the

users to quickly navigate back to previous areas of

interest within the local view.

3.9 Performance

A typical problem with existing ontology visualiza-

tion tools and editors is slow loading times. This is

predominantly due to the OWL standard for ontol-

ogy speciﬁcation being an XML-based language and

the default technique for reading it being via the Java

OWLAPI. We found this to be prohibitively slow for

loading large ontologies (ten to ﬁfteen minutes to read

the entire Galen Ontology and more than thirty min-

utes to read the Gene Ontology). Thus, we believe

that it makes much more sense to use a database to

store the ontology and efﬁciently load sections of it as

required. The database can then be optimized for the

particular queries that individual systems require.

We have developed a translation tool that reads an

OWL ontology and stores its contents in a SQLite

database. Our ontology visualization engine can sub-

sequently read the ontology data from the database

much more quickly. The calculation of key concepts

is also computationally expensive. Hence, the key

concepts are stored in the database to speed up sub-

sequent loading by avoiding recalculation. Table 2

shows some time measurement (in msec.) for loading

and working with various-sized real-world ontologies

from database. It can be observed that for medium-

to large-sized ontologies, our database-based loading

approach (second row for each ontology) achieves a

dramatic performance improvement of up to 40-fold.

http://www.sqlite.org/ (accessed 2012-11-28)

ACKNOWLEDGEMENTS

NICTA is funded by the Australian Government as

represented by the Department of Broadband, Com-

munications and the Digital Economy and the Aus-

tralian Research Council. We acknowledge the sup-

port of the ARC through Discovery Project Grant

DP0987168 and DP110101390.

REFERENCES

Aroyo, L., Welty, C., Alani, H., Taylor, J., Bernstein, A.,

Kagal, L., Noy, N. F., and Blomqvist, E., editors

(2011). The Semantic Web - ISWC 2011 - 10th Inter-

national Semantic Web Conference, Bonn, Germany,

October 23-27, 2011, Proceedings, Part I, volume

7031 of Lecture Notes in Computer Science. Springer.

Ashburner, M., Ball, C. A., Blake, J. A., et al. (2000). Gene

Ontology: Tool for the Uniﬁcation of Biology. Nat

Genet, 25(1):25–29.

Baader, F. and Nutt, W. (2003). Basic description logics. In

Baader, F., Calvanese, D., McGuinness, D., Nardi, D.,

and Patel-Schneider, P., editors, The description logic

handbook: theory, implementation, and applications,

pages 43–95. Cambridge University Press.

Bizer, C., Heath, T., and Berners-Lee, T. (2009). Linked

data — The story so far. International Journal

on Semantic Web and Information Systems (IJSWIS),

5(3):1–22.

Buchheim, C., J

unger, M., and Leipert, S. (2002). Improv-

ing Walker’s algorithm to run in linear time. In Graph

Drawing, pages 347–364. Springer.

da Silva, I. C. S. and Freitas, C. M. D. S. (2011). Using visu-

alization for exploring relationships between concepts

in ontologies. In Proceedings of the 2011 15th In-

ternational Conference on Information Visualisation

(IV’11), pages 317–322. IEEE Computer Society.

Dwyer, T., Koren, Y., and Marriott, K. (2006). Ipsep-

cola: An incremental procedure for separation con-

straint layout of graphs. Visualization and Computer

Graphics, IEEE Transactions on, 12(5):821–828.

Dwyer, T., Marriott, K., and Wybrow, M. (2008). Dun-

nart: A constraint-based network diagram authoring

tool. In Tollis, I. G. and Patrignani, M., editors, Graph

Drawing, volume 5417 of Lecture Notes in Computer

Science, pages 420–431. Springer.

VisualizationofLargeOntologieswithLandmarks

469

Erwig, M. (2003). Xing: A visual xml query language.

Journal of Visual Languages & Computing, 14(1):5–

45.

Grau, B. C., Horrocks, I., Motik, B., Parsia, B., Patel-

Schneider, P., and Sattler, U. (2008). OWL 2: The

next step for OWL. Journal of Web Semantics: Sci-

ence, Services and Agents on the World Wide Web,

6:309–322.

Horrocks, I., Patel-Schneider, P. F., and van Harmelen, F.

(2003). From SH I Q and RDF to OWL: The Making

of a Web Ontology Language. Journal of Web Seman-

tics, 1(1):7–26.

Howse, J., Stapleton, G., Taylor, K., and Chapman, P.

(2011). Visualizing ontologies: A case study. In

(Aroyo et al., 2011), pages 257–272.

Katifori, A., Halatsis, C., Lepouras, G., Vassilakis, C., and

Giannopoulou, E. G. (2007). Ontology visualization

methods - a survey. ACM Comput. Surv., 39(4).

Katifori, A., Torou, E., Halatsis, C., Lepouras, G., and Vas-

silakis, C. (2006). A Comparative Study of Four On-

tology Visualization Techniques in Prot

e: Exper-

iment Setup and Preliminary Results. In Informa-

tion Visualization, 2006. IV 2006. Tenth International

Conference on, pages 417 –423.

Lanzenberger, M., Sampson, J., and Rester, M. (2010). On-

tology visualization: Tools and techniques for visual

representation of semi-structured meta-data. J. UCS,

16(7):1036–1054.

Motta, E., Mulholland, P., Peroni, S., d’Aquin, M., G

omez-

erez, J. M., Mendez, V., and Zablith, F. (2011). A

novel approach to visualizing and navigating ontolo-

gies. In (Aroyo et al., 2011), pages 470–486.

Mukherjea, S. and Hara, Y. (1997). Focus+ context views of

world-wide web nodes. In Proceedings of the eighth

ACM conference on Hypertext, pages 187–196. ACM.

Peroni, S., Motta, E., and d’Aquin, M. (2008). Identifying

key concepts in an ontology, through the integration

of cognitive principles with statistical and topological

measures. In Domingue, J. and Anutariya, C., editors,

ASWC, volume 5367 of Lecture Notes in Computer

Science, pages 242–256. Springer.

Rogers, J. and Rector, A. (1996). The galen ontology. In

Proceedings of Medical Informatics Europe (MIE 96),

pages 174–178. IOS Press.

Ruttenberg, A., Rees, J., Samwald, M., and Marshall, M. S.

(2009). Life Sciences on the Semantic Web: the Neu-

rocommons and Beyond. Brieﬁngs in Bioinformatics,

10(2):193–204.

Smith, B. et al. (2007). The OBO Foundry: Coordinated

Evolution of Ontologies to Support Biomedical Data

Integration. Nat Biotech, 25(11):1251–1255.

Stearns, M. Q., Price, C., Spackman, K. A., and Wang, A. Y.

(2001). SNOMED clinical terms: overview of the de-

velopment process and project status. In Proc AMIA

Symp, pages 662–666, College of American Patholo-

gists, Northﬁeld, IL, USA.

Sugiyama, K., Tagawa, S., and Toda, M. (1981). Methods

for visual understanding of hierarchical system struc-

tures. IEEE Transactions on Systems, Man, and Cy-

bernetics, 11(2):109–125.

Vercruysse, S., Venkatesan, A., and Kuiper, M. (2012).

OLSVis: an animated, interactive visual browser for

bio-ontologies. BMC Bioinformatics, 13(1):116+.

Walker, J. Q. (1990). A node-positioning algorithm for

general trees. Software Practice and Experience,

20(7):685–705.

Wybrow, M., Marriott, K., and Stuckey, P. (2010). Orthogo-

nal connector routing. In Graph Drawing, pages 219–

231. Springer.

IVAPP2013-InternationalConferenceonInformationVisualizationTheoryandApplications

470