Presentation is loading. Please wait.

Presentation is loading. Please wait.

From Hierarchies to Polyarchies: Visualizing Multiple Relationships

Similar presentations


Presentation on theme: "From Hierarchies to Polyarchies: Visualizing Multiple Relationships"— Presentation transcript:

1 From Hierarchies to Polyarchies: Visualizing Multiple Relationships
George G. Robertson Microsoft Research

2 What is the problem? Hierarchies are very common
20 years of hierarchy visualization R&D Significant problems remain New problems appearing (Multiple Hierarchies) Hierarchies (or trees) are one of the most common information structures A typical computer user probably interacts with hierarchies dozens to hundreds of times each day Distinguish between networks, DAG’s, and hierarchies Network – general graph (e.g., the web) DAG – Directed Acyclic Graph - a subset; no cycles Hierarchy – a subset; each node has only one parent Hierarchy visualization is also important because it can be used to view Networks and DAGs with some preprocessing Over the last 20 years there has been a lot of research on how to effectively display and interact with hierarchies In spite of that, we still have not solved some basic problems In addition, as the industry scales to larger enterprise-wide problems, we are beginning to see uses of multiple hierarchies Which current approaches do not handle At Microsoft Research, we set out last Fall to understand current problems with Hierarchy Visualization and propose new solutions I’m going to tell you about of what we have found so far

3 Current Approaches Many 2D and 3D hierarchy visualizations
Each works for some tasks and some scales Very few have had user testing Windows Tree Control Many observed problems Many hierarchy visualization techniques have been proposed over the last 20 years Each solves some problems None solve all problems Note that very few of these have actually been tested with real users I’ll show you a survey of some of these techniques Windows Tree Control is most commonly used hierarchy visualization Many problems are reported by customers and observed in formal usability testing Most frequently requested visualization solution from MS product teams

4 What’s wrong with this picture?
The Windows Tree Control It is used in many different ways: Windows Explorer, file open/close dialogue, History, Favorites, … It does some things well – It can be used in a very small space. But, it has a lot of problems …

5 Problems: Fitting Data
Extreme aspect ratio (broad and shallow) May be multiple hierarchies Scaling issues Two classes of problems: The first is how well the visualization fits the data From our research, we see that most real-world hierarchies tend to be very broad and shallow For example, looking at either file systems or mail folders, we see not many levels, but moderate to high fan-out We are beginning to see problems that involve more than one hierarchy, but most tools only show one. For example, an enterprise may have many databases about people, and want to view them together We know that a given visualization works well within some limits on the number of nodes or levels, But probably does not work well when the data scales beyond those limits

6 Problems: Cognitive Overhead
Loss of context Or loss of detail Separate detail/overview  extra overhead Multiple focus is difficult Which item is open? The second class has to do with the cognitive demands placed on the user by the visualization For any hierarchy more than about 80 nodes, it is very difficult to show all the details on one screen Leads to either a loss of context or a loss of detail A common early solution was to provide two views, one for details and one for an overview But, this forces the user to shift focus of attention from one place to another And can lead to losing context Many user tasks require looking at more than one thing at a time, but most current interfaces do not allow multiple focus Once a node has been selected (or opened), and the user scrolls away from it, the user quickly loses track of where the selected node is

7 Basic View Strategies Two view (separate detail/overview views)
Distorted view Distorted data: fisheye Distorted space: 3D, hyperbolic Focus in Context (integrated view) As we begin to look at the history of visualization techniques, you will see three basic view management strategies Two view – separate detail and overview views Distortion Of the data (e.g., filtering out less interesting material) Of the space Focus in context displays provide a seamless integration of detail and context

8 Basic Visualization Approaches
Indentation Tree control Fisheye Containment Treemaps Pad++ Clustering Galaxy of News ThemeScape Hot Sauce Geographic Floor plans Street maps Node-link diagrams 2D diagrams SemNet Cone Tree Fisheye Cone Tree Hyperbolic viewer FSN XML3D Five basic approaches to visualizing hierarchies Indentation The early tree views used this Fisheye added a distortion technique to increase amount of useful information on screen Containment Treemaps are a 2D space-filling visualization of hierarchies using containment Pad++ is a 2D plus zoom containment system (where some levels are hidden because of zoom factor) Clustering data These techniques are for implicit hierarchies derived from content analysis Geographic Sometimes hierarchy has a direct geographic mapping or analogy Node-link diagrams Probably the most explored in the Information Visualization community Let’s go through some examples of these in chronological order

9 Smalltalk File Browser - 1979
Xerox PARC 1979 One of the earliest hierarchy visualizations was the Smalltalk File Browser in 1979 Four levels of the hierarchy are shown, but only siblings are shown at each level

10 Fisheye Views – 1986 Furnas, Bell Labs report 1981, CHI’86
In the early 80’s, Furnas, at Bell Labs, developed generalized fisheye views A distortion technique that uses a Degree of Interest function to determine what information is shown and what is not shown This example is a Fisheye Calendar focused around a particular date

11 SemNet - 1986 Fairchild, Poltrock, & Furnas, MCC
Semantic network 3D visualization In the mid-80’s, Fairchild, Poltrock & Furnas at MCC developed SemNet A visualization of a semantic network in 3D Made use of fisheye views to manage complexity First 3D network visualization

12 Cone Tree - 1991 Robertson, Mackinlay & Card, Xerox PARC, CHI’91
Limits: 10 levels 1000 nodes Up to 10,000 At CHI’91, I introduced the Cone Tree hierarchy visualization Work done with Jock Mackinlay and Stu Card at Xerox PARC First 3D hierarchy visualization Solved the loss of context problem with a focus and context display Selected node and path to root rotate around to be shown in front So, the context is always easily visible Problems Limits on scale: 10 levels and 1000 nodes for most efficient view But can work up to 10,000 nodes Occlusion forces user to select and animate (essentially a different form of scrolling) Many others have proposed variants that solve these problems to one degree or another

13 Cone Tree – 10,000 nodes Performance problems and Visual clutter
This is what the Cone Tree looks like for 10,000 nodes (directories in a file system) Given the hardware performance at the time, this was too slow

14 Fisheye Cone Tree Adding Fisheye solves problems
Speed problem was addressed simply by adding Fisheye view to Cone Tree The Degree of Interest function took into account recent selections as well as the current selection

15 TreeMap - 1991 Johnson & Shneiderman, U. Maryland, Vis’91
Space filling ~3000 objects MicroLogic’s DiskMapper At about the same time, Johnson and Shneiderman at Univ. of Maryland developed Treemaps Space filling hierarchy visualization that uses containment to represent levels of the hierarchy Deals with about 3000 visible nodes Very good at revealing some properties of the nodes For example, amount of disk space mapped to size of box Also used color to encode additional property (e.g., age) Less good at revealing structural relationships This image is from DiskMapper, a commercial product from MicroLogic

16 FSN - 1992 Tesler, SGI, 1992 2D layout on 3D surface
Loss of context when examining detail Hyper-G (Graz; Andrews; InfoVis’95) In 1992, Joel Tesler at SGI introduced FSN File System Navigator Information Landscape A 2D tree layout on the ground of a 3D environment Selecting a node flies user to that node Resulting in loss of context Did have a second window with an overview In 1995, Andrews at Univ. of Graz adapted FSN for web viewing in a system called Hyper-G

17 Graphical Fisheye - 1992 Sarkar & Brown, DEC SRC, CHI’92
Also a number of 2D network visualizations Graphical Fisheye by Sarkar & Brown at DEC SRC in 1992 Generalization of Furnas fisheye view to 2D graph layout Here the view on the right is the same structure, but focusing on node 48

18 Rubber Sheet - 1993 Sarkar & Snibbe, Brown, UIST’93
Rubber Sheet by Sarkar and Snibbe at Brown in 1993 Generalization of Graphical Fisheye to allow multiple focus

19 Spiral Visualization - 1994
Mackinlay, Robertson, & DeLine, Xerox PARC, UIST’94 Large DAGs Only shows current path Spiral Visualization, by Mackinlay, Robertson & DeLine at PARC Enormous hierarchies and DAGs Calendar and decision tree Note that only the current path through the DAG is shown

20 Hyperbolic Browser - 1994 Lamping & Rao, Xerox PARC, UIST’94
Projected onto circle 1000’s of nodes Hyperbolic Browser in 1994 by Lamping & Rao at PARC A reaction to occlusion problems in the Cone Tree Use of hyperbolic space instead of 3D perspective Hyperbolic layout projected onto 2D circle Can display 1000’s of nodes, with details diminishing as you near the edge Has property that much information can be packed into the periphery

21 Continuous Zoom - 1995 Bartram et al, Simon Fraser, UIST’95)
Continuous Zoom, by Bartram et al at Simon Fraser Univ. in 1995 Generalization of Rubber Sheet visualization to allow smooth transitions while changing one or more foci of attention

22 Fsviz - 1995 Carriere and Kazman, Waterloo, InfoVis’95
Up to nodes with no occlusion Fsviz, by Carriere and Kazman at Waterloo Univ. in 1995 Addressed occlusion problem in Cone Tree by moving to different viewpoint Can layout 5000 nodes with no occlusion But, they are small and far away

23 Butterfly - 1995 Mackinlay, Xerox PARC, CHI’95
Butterfly, by Mackinlay at PARC in 1995 citation index Enormous DAG Deals with asynchronous searches Structure evolves

24 Index View - 1995 T. Masui, Sharp, UIST95 Vertical picks focus
Horizontal controls zoom Aaaamon.dll Edb500.dll Label.exe Oakley.dll Tapi.dll Wavemsp.dll Ksproxy.ax Ksqmf.ax Ksuser.dll Ksvpintf.ax Lanman.drv Legacy.inf Linkinfo.dill Lmrt.dll Aaaamon.dll Edb500.dll Label.exe Oakley.dll Tapi.dll Wavemsp.dll Ksproxy.ax Ksuser.dll Legacy.inf Lmrt.dll Aaaamon.dll Edb500.dll Label.exe Oakley.dll Tapi.dll Wavemsp.dll Ksproxy.ax Lmrt.dll Aaaamon.dll Edb500.dll Label.exe Oakley.dll Tapi.dll Wavemsp.dll Index View, by Masui at Sharp in 1995 Primarily a list viewer for a small space, but can also be used for hierarchies User moves mouse vertically to pick focus Then horizontal to control zoom level Does have a discoverability problem So only works well if user knows what they are looking for Partially addressed by using a permuted index

25 H3 - 1997 Munzner, Stanford Univ., InfoVis’97 Projected onto sphere
20,000 nodes H3 layout algorithm, by Munzner at Stanford in 1997 An attempt to merge ideas from Hyperbolic browsers and Cone Trees Projection is onto the surface of a sphere with the focus node at the center of the sphere Able to display 20,000 nodes

26 Reconfigurable Disk Tree - 1998
Jeong and Pang, UC Santa Cruz, InfoVis’98 Number of identifiable nodes: – 4000 Reconfigurable Disk Trees, by Jeong and Pang at UC Santa Cruz in 1998 Another attempt to get around occlusion problems in Cone Tree Both by using a different view point And by collapsing the cones into disks Displays 3000 – 4000 identifiable nodes

27 Disk Tree - 1998 Chi et al, Xerox PARC, CHI’98
Compact 2D representation Disk Tree, by Chi et al at PARC in 1998 Compact 2D representation uses concentric circles, with a hierarchy level for each successive circle Used to show structure of a web site Line size and brightness encode page access frequency Color encodes page lifecycle stage Advantages No occlusion Compact Pattern can be compared over time to observe changes

28 Sunburst - 2000 Stasko & Zhang, Georgia Tech, InfoVis 2000
Radial space-filling Techniques for viewing more detail Sunburst visualization, by John Stasko and Eugene Zhang at Georgia Tech, presented at InfoVis 2000 Basic Sunburst is similar to Disk Tree, except each level is space-filling Study compared Sunburst to Treemap Subjects preferred Sunburst because of more explicit representation of structure Focus of InfoVis work was on new techniques to show additional detail Angular detail of a focus region “Detail outside” shrinks context to inner space and expands focus region to outer circles “Detail inside” is opposite

29 Visualization Taxonomy - 1994
Noik (Graphics Interface’94) Implicit (use of perspective) Continuous focus and context Filtered (removing items of low interest) Discrete focus and context Distorted (size, shape, position of elements) Adorned (color, texture) - Noik’s taxonomy sheds light on what these various techniques are doing - Implicit (use of perspective) - Advantage: continuous focus and context - Filtered (removing items of low interest) - Fisheye and dynamic queries - Disadvantage: discrete focus and context - Distorted (size, shape, position of elements) - Fisheye also does this - Can make it hard for user to track items of interest - Adorned (color, texture)

30 Unresolved Problems Multiple focus
We completed a whirlwind survey of many hierarchy visualization techniques developed over the last 20 years Each solves some of the problems we originally observed None solve all the problems, and there are still some problems that need more work For example, multiple focus has been addressed by only a few of the techniques Yet, in real use, this is one of the biggest problems Here are some mockups of some ideas we have been playing with

31 Folding Paper - proposal
This uses a metaphor of folding paper Areas of interest are stretched out Other areas are stacked A focus area can have a second view that is magnified to reveal some detail

32 Multiple Focus in 3D - proposal
This is a similar idea, but uses 3D to display a series of columns, one for each level When there are areas of multiple focus, more than one list will appear in a column 3D is used only to reduce screen space, not to encode any properties of the hierarchy

33 InfoBowl - proposal The InfoBowl is a suggestion for how to introduce multiple focus into hyperbolic browsers Each focus has a second level hyperbolic browser for exploring the next level of detail Note: These are meant to be brainstorming examples. The hope is that more people will explore the problem of multiple focus in hierarchy visualizations

34 Unresolved Problems Multiple hierarchies Multiple focus
Almost all visualization techniques so far have shown only one hierarchy But, we are beginning to see examples where the user must deal with and understand multiple hierarchies

35 Multiple Hierarchies One hierarchy changing over time Time Tube
Taxonomy visualization MultiTrees (shared subtrees) XML3D Polyarchy (multiple intersecting hierarchies) There are at least three kinds of multiple hierarchy problem Two recent projects have explored what it means to view a hierarchy that changes over time In essence, these are looking at multiple hierarchies that are very similar MultiTrees are multiple hierarchies with shared subtrees The third kind, which we call Polyarchies, are multiple intersecting hierarchies

36 Time Tube - 1998 Chi et al., Xerox PARC, CHI’98
In 1998, Chi et al at PARC introduced the Time Tube A sequence of Disk Trees used to visualize changes in a web site over time This is an example of viewing multiple similar hierarchies

37 Taxonomy Visualization - 2000
Graham et al., Napier Univ. IJHCS 2000 Recently, Graham et al at Napier Univ. explored visualizing changing biological taxonomies In this example, individual objects are tracked across the hierarchies Graham told me informally that this approach was abandoned when they tried to scale it up to real sized problems

38 MultiTrees - 1994 Furnas & Zacks, Bell Core, CHI’94
DAG’s with shared subtrees MultiTrees were introduced by Furnas and Zacks at Bell Core in 1994 They are basically large DAG’s with shared subtrees The typical example is about a class syllabus created by one professor A second professor reuses some of the material, but only a subset A third professor uses a different subset, possibly reusing some of the material introduced by the second professor

39 XML3D - 2000 Munzner, Stanford, IJHCS 2000
Recently, Munzner at Stanford used the H3 hyperbolic projection onto a sphere To encode information from XML structures with multiple inheritance This is essentially a MultiTree problem User study 12,000 node topic hierarchy (Snap.com) XML3D vs web-based solution vs Windows tree control XML3D was reliably faster than the two 2D solutions

40 Polyarchy Visualization Problem: People and Resources Example
Multiple Hierarchies Exist Direct reporting Cost or Profit Center Location Implicit relationships MultiTrees involve multiple hierarchies pointing to the same set of leaves (I.e., shared subtrees) Polyarchies are about multiple intersecting hierarchies (a superset of MultiTrees) He are some examples from a particular problem we are currently working on. Problem comes from Metadirectory Services product How do you bring together various information sources in an enterprise and provide a common access point? There will be multiple hierarchies and other relationships In this example, we are looking at information about people from four different information sources In spite of having multiple hierarchies, the current solution one shows one hierarchy But only one hierarchy is shown

41 Show Other Hierarchies
Find: Daniel Robbins Naming hierarchy is default Allow maintainer to define hierarchies Show search results in current view Morph between views There is a default hierarchy, called the naming hierarchy We give the metaverse designer a way to specify what hierarchies should be exposed to the user Search results are always displayed in the currently viewed hierarchy We morph between views – we call this a visual pivot

42 Visualization Opportunities
Perceptual cues to provide pattern information User can search or browse Show hierarchies an object participates in Show relationships between hierarchies Show relationships between objects Given the existence of multiple intersecting hierarchies and other relationships How do we display this complexity to the user in a way that makes sense? This is work we are currently engaged in In particular, we want to show the hierarchies an object participates in And the relationships between those hierarchies As well as the relationships between the objects in the hierarchies

43 One Selection, One Hierarchy
Here is what the current prototype looks like Upper left: user enters search terms and identifies their type Lower left: results of search are displayed Upper right: user picks a dimension (a hierarchy or relationship) Lower right: the current selected people are displayed in the current selected dimension In this case, Anoop Gupta is shown in the management hierarchy

44 Visual Pivot – Other Hierarchies
The user can pivot to any other dimension that the selected node participates in This movie shows the original visual pivot animation First pivoting to the business unit hierarchy Then, related people Finally, office location

45 Relationships Between People
The Related People dimension is particularly interesting It shows a complex relationship Management chain Direct reports Immediate peers Some dimensions are simple, others quite complex

46 Relationships Between Multiple People
Often, users want to see how various people relate to each other The interface allows the user to add as many people to the display as desired Here we see two people in the management hierarchy It finds the manager they have in common and displays both paths Notice in all of these displays, we are managing an enormous database by showing the minimum set of information that meets the user’s needs Allowing the user to easily change the information shown

47 Unresolved Problems Multiple focus Multiple hierarchies
Evaluate hierarchy visualizations Another key unresolved problem is that we must evaluate these techniques Identify problems that can be resolved in later versions Demonstrate the actual value for an end user Basis for economic and managerial decisions for investment in this technology

48 Polyarchy Visualization User Studies
Study 1: Mockup of visual pivot Issues list guided development of prototype Study 2: Prototype: 2D vs 3D Visual Pivot animation was misleading Animation Speeds were too slow To illustrate, let’s look at the studies done on the Polyarchy Visualization We have done 4 user studies so far First, we built a mockup of the visual pivot and did a simple usability study to determine if users could make sense of the complexity Users had no difficulty in understanding what visual pivot was about Users had high satisfaction ratings of the 3D visual pivot The issues list then guided the development of a working prototype Second, we did a usability study of a working prototype We hoped to get performance data comparing 2D and 3D solutions But, there were usability issues that still needed to be resolved The visual pivot animation caused some problems because of poor readability of text during the pivot The animation speeds were clearly too slow

49 Polyarchy Visualization User Studies
Study 3: Animation Styles and Speeds Six animation styles: Picked two best Twice as fast as study 2: Still too slow Study 4: Prototype: 2D vs 3D Identified most effective animation style Identified best speed range The third study was specifically designed to evaluate six alternative animation styles and to identify the appropriate speed range The most recent study was another performance study, similar to Study 2 except that the usability issues were resolved

50 Two Styles of Visual Pivot
Sliding Rotating The results of the third study indicated that two particular pivot styles should be further evaluated Initial pivots had a final step that moved to a preferred position and scale The study showed that people had trouble understanding that because the new information was not always readily visible The revised animations do that movement first rather than last The study also showed that users had difficulty reading text while rotating around the vertical axis Revised pivot is around the horizontal axis A new pivot was added that slides the new information from right to left, then up a small amount This allows reading the text of old and new side by side

51 Visual Pivot Styles Sliding versus Rotating
The fourth study tested these styles Looking at performance data, there is a reliable advantage to the sliding pivot Satisfaction data was mixed, with no clear preference

52 Visual Pivot Styles Learning Effects
Performance data shows a reliable learning effect, with the sliding pivot easier to learn

53 Visual Pivot Speed Pivot times are broken into three phases:
Reposition Pause after new information is shown Pivot Total Times tested: 0 sec, ½ sec, 1 sec, 2 sec User task completion time was reliably slower with the slow animation But, the other animation speeds, as well as no animation, were not reliably different Can conclude that 3D animation does not incur any performance penalty Preference data indicated a reliable preference for the “fast” animation We are continuing the study of these visual pivot techniques

54 Unresolved Problems Get the solutions out to real users!
Multiple focus Multiple hierarchies Evaluate hierarchy visualizations Get the solutions out to real users! And finally, we need to get hierarchy visualization solutions demonstrated to be of value out into the mass market

55 Summary Use perception to reduce cognitive overhead
Need continued research to solve problems Help user focus on multiple items of interest without losing context Multiple hierarchies and other relationships Handle larger scale Evaluate  Iterate  get results out In summary We need to continue exploring ways to use perception to reduce cognitive overhead We need to continue research into novel hierarchy visualizations We need better ways to help the user focus on multiple items of interest without losing context We need new solutions for multiple hierarchy and polyarchy visualization Handle larger scale – 100,000’s to millions And finally, we need more user studies to evaluate what we are doing This should lead to better designs And is ultimately a requirement for getting our results out to the mass market


Download ppt "From Hierarchies to Polyarchies: Visualizing Multiple Relationships"

Similar presentations


Ads by Google