Interactive Visualization of Large Graphs and Networks

1
Interactive Visualization of Large Graphs and
Networks

• Tamara Munzner
• Stanford University
• Computer Science Department

2
Contributions

• analysis of three software systems
• relating intended tasks to spatial layout, visual
  encoding choices
• two novel layout/drawing algorithms
• scalable
• targeted

3
Three Visualization Systems

• general

domain specific
graph drawing
infovis
H3
PM
Const

• Constellation
• parsed dictionaries
• refine algorithms

• H3
• web hyperlinks
• quasi-hierarchical

• Planet Multicast
• MBone tunnels
• find badly placed

4
Talk Outline

• graph drawing, information visualization
  background
• software systems
• goal
• previous work
• video
• discussion
• evaluation
• general discussion
• conclusion

5
Graph Drawing

• automatic layout and drawing of node-link graphs

6
Goal help humans understand

• aesthetic criteria
• minimize crossings
• expose structure hierarchy, symmetry, circular

Tom Sawyer Software. Hierarchical Toolkit
Tom Sawyer Software. Symmetric Toolkit
Tom Sawyer Software. Circular Toolkit
7
System Scalability, Data Set Size
previous systems
H3
data sets
my systems
Planet Multicast
mid-size web sites
Constellation
Web (pages)
exceptional GD systems (dot, Gem3D)
MBone (tunnels)
Stanford graphics site
most GD systems
my site
Net (routers)
Net (hosts)
dictionary
GEB figure
manual
node count, log scale
8
Fundamental Idea

• extend reach of graph drawing with information
  visualization approach
• techniques
• interactivity
• incorporate domain-specific information

9
Information Visualization

• external visual representation of data, exploits
  perceptual system to reduce human cognitive load
• find appropriate visual metaphor for data that is
  not implicitly spatial

10
Interactivity

• mimic reality
• 2D paper pan, zoom
• 3D object rotate, translate, scale
• beyond
• semantics impossible in real world
• distortion, multi-scale

11
Domain/Task Focus

• user-centered design, ethnography
• understand high level goals
• maintain web site
• break down into lower level tasks
• minimize user navigation to important pages
• find and fix broken links
• design visual encoding
• evaluate effectiveness

12
Evaluating Visualization Systems

• quantitative algorithmic improvements
• conceptual framework analysis
• impact/adoption
• user studies
• anecdotal evidence

13
System 1 H3

• time 1996-8
• data web hyperlinks
• quasi-hierarchical graphs can find reasonable
  spanning tree using domain-specific information
• goal scalability
• method 3D hyperbolic

14
Background Hyperbolic Space

• FocusContext distortion
• project from infinite hyperbolic to finite
  euclidean
• pick best model for useful distortion

conformal geodesics warped
projective angles warped
4x4 matrix
15
Background Hyperbolic Space

• exponential room in space
• exponential number of tree nodes

2D hyperbolic plane
hyperbolic hemisphere area exponential 2p sinh r
2
euclidean hemisphere area geometric 2pr
2
Thurston and Weeks, The Mathematics of Three
Dimensional Manifolds, Scientific American
16
Previous Work Hierarchies

• hierarchies

• Cone Trees
• Robertson, Mackinlay, Card 91

Tree Maps Johnson, Shneiderman 91
Robertson, Mackinlay and Card. "Cone Trees
Animated 3D visualizations of hierarchical
information.
Johnson and Shneiderman. Treemaps A
Space-filling Approach to the Visualization of
Hierarchical Information

• distortion Furnas, Brown, Carpendale, Keahey

17
Previous Work Distortion Hierarchy

• 2D Hyperbolic Tree
• Lamping, Rao, Pirolli 94,95
• scalability analysis later
• Fractal Koike, Yoshihara 93
• SHriMP Storey, Muller 95
• dont scale
• taxonomy Noik 94

Lamping, Rao, and Pirolli. A FocusContent
Technique Based on Hyperbolic Geometry for
Viewing Large Hierarchies.
18
(No Transcript)
19
H3 Layout

• novel layout algorithm detailed in thesis
• hemisphere surface instead of linear
  circumference
• bottom-up pass compute hemisphere sizes
• top-down pass place child on parent surface

20
Information Density Scale
Lamping, Rao, and Pirolli. A FocusContent
Technique Based on Hyperbolic Geometry for
Viewing Large Hierarchies.
21
Information Density Codimension

• want balance between clutter and void
• topological approach to describing density
• difference between structure and surrounding space

sparse
dense
Carpendale, Cowperthwaite, and Fracchia.
Extending Distortion Viewing from 3D to 2D.
22
Evaluation Scalability

• drawing constant
• incremental
• exception precision
• layout linear in E
• 110,000 edges in 12 seconds given DFS input
• limits
• computational global layout in main memory
• cognitive disorientation past 100K nodes
• large neighborhood not global overview
• future landmarks, LOD, abstraction

23
Evaluation Impact

• product from SGI
• Site Manager aimed at web content creators
• bundled starting with Irix 6.3
• research use of library
• interface for Skitter Internet tomography data
• analysis of Autonomous System data
• viewer use
• 6 researchers converted data to use viewer
• image use
• 6 reprint requests

24
Evaluation User Study

• Risden, Czerwinski, Munzner, Cook 00
• compared 3 browsers for adding content to
  collection of web pages

snap portal (Yahoo style)
XML3D H3 lists
collapsible tree
25
User Study Results

• reliably faster for existing category task
• no decline in quality for new category task

differences statistically significant
differences statistically insignificant
26
System 2 Planet Multicast

• time 1996
• joint work
• Hoffman, Claffy, Fenner
• data MBone tunnels
• task find badly placed tunnels
• goal simple baseline
• method 3D geographic

27
Previous Work Geographic Network

• SeeNet3D
• Cox, Eick 95
• arcs on globe layout

Cox and Eick. 3D Displays of Network Traffic.

• SeeNet
• Becker, Eick, Wilks 95

• NSFNet
• Cox, Patterson 92

Becker, Eick, and Wilks. Visualizing Network Data
Cox and Patterson. Visualization Study of the
NSFNet.
28
(No Transcript)
29
Geographic Layout

• distance as stand-in for resource usage
• partially correlated
• geographical determination arduous
• major scalability problem
• immediate comprehension
• evocative, many image reprints
• Wired, National Geographic
• still picture captures much of function

30
Evaluation Anecdotal Insights

• gt pen-mbone-1.sprintlink.net(204.213.238.11)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/64/tunnel
• gt elm.can.net(199.246.170.7) dc-mbone-1.sprintlink
  .net(206.229.87.99) 1/64/tunnel
• gt boston.terra.net(199.103.128.254)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/0/tunnel/querier
• gt NS.FLSIG.ORG(192.153.117.162)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/64/tunnel
• gt ace.mid.net(198.247.225.251) dc-mbone-1.sprintli
  nk.net(206.229.87.99) 1/64/tunnel
• gt fw-mbone-1.sprintlink.net(206.61.106.99)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/16/tunnel
• gt gateway10.crawford.com(198.69.210.2)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/32/tunnel
• gt csce-2--rngm-nb-f-1.net.tamu.edu(128.194.1.11)
  dc-mbone-1.sprintlink.net(206.229.87.99)
  1/64/tunnel
• ...

31
System 3 Constellation

• time 1998-9
• joint work
• Guimbretière
• data MindNet query results
• task plausibility checking for linguists
• method 2D custom
• goal targeted

32
Definition Graph

• dictionary entry sentence
• nodes word senses
• links relation types

33
Semantic Network

• definition graphs as building blocks
• unify shared words
• large network
• millions of nodes
• grammar checking now, translation future
• global structure known dense
• probes return local info

34
Path Query

• best N paths between two words
• words on path itself

• definition graphs used in computation

35
Task Plausibility Checking

• paths ordered by computed plausibility
• researcher hand-checks results
• high-ranking paths believable?
• believable paths high-ranked?
• stop words

36
Top 10 Paths kangaroo - tail
37
Goal

• create unified view of relationships between
  paths and definition graphs
• shared words are key
• thousands of words (not millions)
• special-purpose algorithm debugging tool
• not understand the structure of English

38
Previous Work Semantic Networks

• SemNet
• Fairchild, Poltrock, Furnas 88
• multiple 3D layouts

• Visual Thesaurus
• Thinkmap applet
• casual browsing, constant motion
• lt 20 nodes

Fairchild, Poltrock, and Furnas. SemNet
Three-Dimensional Graphic Representations of
Large Knowledge Bases.
Thinkmap applet. www.thinkmap.com cited 3/09/00.
39
(No Transcript)
40
Traditional Layout

• avoid crossings
• reason avoid false attachments

A
A
B
B
C
D
C
artifact salience
ambiguity
41
Information Visualization Approach

• spatial position is strongest perceptual cue
• encode domain specific attribute
• plausibility gradient

42
Constellation Semantic Layout

• novel layout algorithm detailed in thesis
• paths as backbone, definition graphs attached
• curvilinear grid
• iterative design for maximum semantics with
  reasonable information density
• allow crossings for long-distance proxy links

43
Selective Emphasis

• highlight sets of boxes and edges
• interaction
• additional perceptual channels
• avoid perception of false attachments

44
Evaluation Layout Effectivness
45
Evaluation Layout Comparison
46
Talk Outline

• graph drawing background
• software systems
• goal
• previous work
• video
• discussion
• evaluation
• general discussion
• conclusion

47
Visual Salience

• Planet Multicast
• long-distance tunnels
• H3
• distant points of possible interest
• fringe aggregate information
• Constellation
• selective emphasis
• word size tied to importance

48
Canonical Word Size
49
Hidden State

• Constellation avoids hidden state
• change salience instead of toggle drawing
• closed world assumption
• if not visible, doesnt exist
• easy to forget previous actions
• false negative conclusions
• H3, PM do have hidden state
• non-tree links sometimes drawn
• intra-city tunnels never drawn

50
Graph Functions

• structure discovery
• pure spatial layout
• implicit in traditional graph drawing
• contextual backdrop
• linked view

51
Graph Functions

• structure discovery
• contextual backdrop
• additional visual encoding
• color, linewidth, shape, enclosure
• combination more than sum of parts
• linked view

52
Contextual Backdrop
53
Graph Functions

• structure discovery
• contextual backdrop
• linked view
• brushing Becker and Cleveland 88
• invoke other software components

54
Linked View
55
Contributions

• detailed analysis of three software systems
• interactive, range of domain specificity
• relate intended tasks to spatial layout, visual
  encoding
• two novel layout/drawing algorithms
• Constellation
• targeted design
• H3
• scales 100x beyond previous work
• product, user study

56
(No Transcript)