CAR-TR-772 March 1995
Users often must browse hierarchies with thousands of nodes in
search of those that best match their information needs. The PDQ
Tree-browser (Pruning with Dynamic Queries)
visualization tool was specified, designed and developed for this
purpose. This tool presents trees in two tightly-coupled views,
one a detailed view and the other an overview. Users can use dynamic
queries, a method for rapidly filtering data, to filter nodes
at each level of the tree. The dynamic query panels are user-customizable.
Subtrees of unselected nodes are pruned out, leading to compact
views of relevant nodes. Usability testing of the PDQ Tree-browser,
done with 8 subjects, helped assess strengths and identify possible
improvements. The PDQ Tree-browser was used in Network Management
(600 nodes) and UniversityFinder (1100 nodes) applications. A
controlled experiment, with 24 subjects, showed that pruning significantly
improved performance speed and subjective user satisfaction. Future
research directions are suggested.
please address correspondence to Ben Shneiderman
To appear in International Journal of Human-Computer Studies
Decisions are an integral part of human life. Whether deciding what movie to see or choosing which universities to apply to, people are constantly faced with decisions. Many everyday decisions and in engineering applications require the selection of one or a few elements from many, possibly thousands of elements. In such cases, users often try to make a ``good'' choice by deciding first what they do not want, i.e. they first try to reduce the data set to a smaller, more manageable size. After some iterations, it is easier to make the final selection(s) from the reduced data set. This iterative refinement or progressive querying of data sets is sometimes known as hierarchical decision-making.
A hierarchical data set (HDS) (also called a ``tree'') organizes data points into a hierarchy. HDSs are common, e.g. sales & budget data, catalogs of products, library indices, computer file systems etc. A HDS can be filtered (queried and reduced) effectively using a hierarchical-decision making process because the data is inherently hierarchical.
The original motivation for this work was our research on user
interfaces for network management (Kumar, Plaisant, Teittinen,
& Shneiderman, 1994). While working on a user interface for
satellite network configuration, we were faced with the task of
choosing one leaf node from a large tree of thousands of nodes.
Further, the task was such that the number of interesting leaf
nodes could be reduced drastically based on selection criteria
at various levels in the tree. This problem prompted us to design
and implement the PDQ (Pruning with Dynamic Queries)
2 PDQ Tree-browser Requirements
Based on our task analyses in the network management scenario and other tree-browsing applications, we specified the requirements for the PDQ Tree-browser visualization tool:
3 Previous Work
Information search is a vast topic (Marchionini, 1995), but previous work on hierarchical data is more limited. Early work emphasized aesthetic and easy-to-read tree and graph layouts (Battista, Eades, Tamassia, & Tollis, 1989), and browsing / exploring / searching trees and graphs, but none of them provide a good solution to the attributes-based querying and browsing problem. Many visual browsers address 2D-browsing, and some allow for manual pruning of uninteresting subtrees / subgraphs. Commonly used tree visualizations include 2D node-link diagrams, space-filling treemaps, 3D node-link diagrams and tables-of-contents or outliners.
An advantage of node-link visualizations of trees is that they are a familiar mapping of structured relationships and therefore easy to understand. They can also display attributes of links by color or size if required. However, node-link diagrams make inefficient use of screen space, and even trees of medium size need multiple screens to be completely displayed. This necessitates scrolling of the diagram and global context is lost, since only a part of the diagram is visible at any given time.
Beard & Walker (1990) used a map window - a miniature of the entire information space with a wire-frame box to aid users in remembering their location. The map window is better known as an overview , the entire information space shown in full size is the Detailed View and the wire-frame box is the field-of-view or the panner (Plaisant, Carr, & Shneiderman, 1995). The field-of-view can be dragged around in the overview to pan the detailed view. Similarly, scrolling the detailed view updates the position of the field-of-view in the overview. Hence, the overview and the detailed view are said to be Tightly-coupled. Beard & Walker found that an overview significantly improves user performance, and the pan technique and the zoom and pan technique were significantly faster than the scroll technique.
Plaisant et al. provide a taxonomy and guidelines for image-browsers. This work attempts to standardize some of the terms being used by researchers today, e.g. Detail View, Field-of-view, etc. Different kinds of browsers such as ``Detail only'', ``One window with zoom and replace'', and ``Tiled multilevel browser'' are shown. The authors identify 5 classes of tasks that are accomplished with image browsers, e.g. open ended exploration, navigation, monitoring, etc.
The treemap visualization of tree structures uses a two-dimensional (2D) space-filling approach in which each node is a rectangle whose area is proportional to some attribute such as node size (Shneiderman, 1992). On the other hand, the treemap algorithm utilizes 100 % of the designated space. Sections of the hierarchy containing more important information can be allocated more display space while portions of the hierarchy which are less important to the specific task at hand can be allocated less space. Treemaps have been used for file management, network management (Kumar et al., 1994), budgets, sports data, etc.
Robertson, Mackinlay, & Card (1991) and Chignell, Zubrec, & Poblete (1993) used 3D node-link diagrams in order to visualize tree structures. Robertson et al., in the Information Visualizer project at Xerox PARC, developed a tool called Cone Trees that allows for animation of 3D trees. They contend that interactive animation can effectively shift cognitive processing load to the perceptual system. They describe gardening operations' (Robertson et al., 1991) where the user can manually prune and grow the view of the tree. Prune and grow operations are done either by menu or by gestures directed at a node. While these gardening operations do help in managing and understanding large, complex hierarchies, this manual mechanism of pruning or growing one subtree at a time, is clearly not sufficient nor very effective in specifying complex searches spanning several levels of the tree.
Chignell et al. (1993) built the Info-TV tool, which allows two styles of pruning:
The authors however, do not describe how the nodes whose sub-trees are to be pruned are specified by the user. It is assumed that the user makes these specifications manually by selecting these nodes.
Visualizations have also been used traditionally to browse of graph structures. Examples of graph visualizations include hypertext graphs, finite-state diagrams, flow-charts, parse-trees, pert-charts, dataflow diagrams, wiring diagrams and hierarchical decompositions.
According to Henry (1992), and Henry and Hudson (1991), the ``best layout'' depends on the user's current region of interest. Consequently, a single layout algorithm cannot always produce the best results. Thus, the ability of users to customize the layout to meet their current needs and interests is essential. Therefore, users must be provided with interactive tools to iteratively dissect large graphs into manageable pieces. Henry and Hudson describe manual selection in which users select nodes and edges using simple direct manipulation techniques, and algorithmic selection in which users apply an algorithm to the graph. But by classifying manual selection into only individual selection or marquee selection, the authors are greatly restricting the range of interesting selection subsets that can be specified by the user. They conclude by identifying two primary future directions: using domain specific graph semantics to guide the layout and the selection, and creating a methodology that can be used to build an interactive system for non-programmers to specify selection and layout algorithms. Our work extends their strategies.
Gedye (1988) built a system that presents users with a list of all the objects. They can choose an object, and then upon choosing one particular relationship, a directed acyclic graph, tree, or an equivalence set is created in a new window. So, the tangled web that would have resulted by showing all the relationships is eliminated by using one window per relationship. Noting that both zooming and panning were inadequate for arbitrary graph structures, the authors implemented pruning in order to assist in the browsing. A subgraph is selected that contains few enough nodes to comfortably fit in the window, and this subgraph is displayed in its entirety. The subgraph obtained by the pruning procedure is called the "display graph".
Schaffer, Zuo, Bartum, Dill, Dubs, Greenberg, & Roseman (1996) found that users performed better using fisheye views of hierarchically-clustered graphs than using full-zoom views. Hollands, Carey, Matthews, & McCann (1989), however found users getting somewhat disoriented while using fisheye views for complex tasks. None of the fisheye view implementations described above allow for attributes-based specification of the foci of interest.
The review of the literature pertaining to tree and graph browsers did not provide a good answer to the problem of specifying selection subsets on large connected data sets like trees and graphs. Those that do allow for uninteresting nodes / subtrees / subgraphs to be pruned out, require users to make subset specifications manually. We believe that Dynamic Queries can be used effectively for this purpose. Dynamic Queries describes the interactive user control of visual query parameters that generates a rapid (100 msec update) animated visual display of database search results. Dynamic queries are an application of the direct manipulation principles in the database environment. They depend on presenting a visual overview, powerful filtering tools, continuous visual display of information, pointing rather than typing, and rapid, incremental, and reversible control of the query (Shneiderman, 1992). Dynamic queries have been applied to browser databases of houses, movies, chemical tables of elements etc. These concepts of dynamic querying and tight-coupling are similar to those of Focusing and Linking (Buja, McDonald, Michalak, & Stuetzle, 1991).
4 PDQ Tree-browser design
Our PDQ Tree-browser design consists of the following features:
4.1.1 Dynamic Queries on Hierarchical Data Sets: Dynamic queries have been applied to data sets consisting of independent data points (Ahlberg and Shneiderman, 1994). In such cases, whether a data point satisfies a given query or not does not affect the outcome for other data points. This is because there are no interrelationships between the data points. Thus, the data set can be thought of as ``flat''. Queries are merely queries on the attributes of individual data points. For example, in the FilmFinder, each movie is an independent data point. Similarly, in the HomeFinder (Shneiderman, 1994), each house is an independent data point.
In a "non-flat'' data set, on the other hand, there are predefined
interrelationships between data points. For example, in a HDS,
some nodes are related to some others by the parent-child relationship
(Kumar et al., 1994). Therefore, whether a node matches
a given query or not might affect some other nodes. Specifically,
while searching a hierarchical data set in a top-down manner (i.e.
parent first), it makes sense to prune out all descendant nodes
of nodes that do not match the query. For example, if users are
looking for departments in universities with low tuition, it makes
sense to eliminate those departments whose universities have high
tuition. When criteria (like low tuition, high average SAT scores,
high placement indices etc.) exist at all or most levels in the
hierarchy, stepwise refinement of the query can be done
to progressively reduce the initial (large) data set into a smaller
set, from which good choices may be made.
4.1.2 Dynamic Query Environments: The HomeFinder and the FilmFinder are examples of systems that provide hard-coded graphical widgets for the user to manipulate in order to dynamically update the visual display. If the user wanted to find homes that were within 2 miles of any hospital, the current HomeFinder interface would have to be reprogrammed.
Therefore, a Dynamic Query Environment , first implemented
in a browser for the National Center for Health Statistics (Shneiderman,
1994), should allow users to customize dynamic query control panels
based on current interests. Users should be able to select what
combination of attributes they wish to query on, and have the
appropriate widgets created, at run-time. The method of selecting
the attributes and creating widgets should be easy, and also allow
for modifications / backtracking. Having the interface not be
application-dependent would have the added advantage of reusability
4.2 Description of the Interface
The PDQ Tree-browser interface consists of two main parts (Figure
The Treeview and Range-slider widgets of the University of Maryland Widget Library TM (Carr, Jog, Kumar, Teittinen, & Ahlberg, 1994) were used in the PDQ Tree-browser implementation.
If users manipulate a widget at the current lowest level displayed, the nodes matching the query at that level are colored yellow, otherwise they are gray. These updates of the data display are real-time (within 100 msec of updates to the control widgets) in accordance with the principle of dynamic queries. Buffering was done in order to make the updates as smooth and flicker-free as possible.
If users manipulate a widget at a level other than the current lowest level, the tree visualization first ``jumps'' to that level, i.e. the level of the widget is made the tree's current lowest level. This is done so that the structure of the tree during direct manipulation of the widgets remains constant and only the colors of the nodes change. Then the nodes at the new lowest level are updated (by coloring yellow or gray) in real-time to show whether they match the query or not.
The tree structure changes only when the current lowest level of the tree is changed, which can be accomplished by users in three ways:
When the current lowest level is changed so as to show more levels, pruning of the tree is done so as to eliminate subtrees of nodes that do not match the query at their own levels. For example, if the user manipulates the range slider for Network ID such that Nets 1 and 2 do not match the query (Figure 2), and then increases the lowest level displayed by 1, the children of nets 1 and 2 are not shown, while children of other nets are shown (Figure 3). Nets 1 and 2 now appear in gray, while all the other nets appear in orange, simply to show that those nodes did match the query at their own level. This feedback enables users to go back and change queries at higher levels, and thus iteratively refine the selection subset.
As explained above, nodes at the current lowest level are colored either yellow or gray, while nodes at higher levels are colored either orange or gray. Also, the level button corresponding to the current lowest level is colored yellow so as to focus the attention of the user to that level. Buttons at other levels are colored gray.
When the current lowest level is changed so as to show fewer levels, the tree is simply ``folded'' back to the new lowest level, and then the nodes at that level are colored either yellow or gray based on the query at that level. Thus, users can easily jump back and forth between levels in order to fine-tune their search.
The Attributes List on the left shows only the attributes corresponding to the current lowest level. Users can access names of attributes at other levels by choosing the appropriate level name (e.g. Network, DPC, etc.) from the attributes menu just above the Attributes List (Figure 1).
There are four feedback indicators, one corresponding to each level (other than the root level), that are updated in order to show the number and proportion of nodes that currently match the query (i.e. number of ``hits''). The proportion of the feedback indicator that is colored yellow corresponds to the proportion of hits. This proportion is a percentage of the total number of nodes at that level, not of the number of nodes at that level currently displayed. The actual number of hits is also displayed at the top of each feedback indicator. Proportion indicators are displayed from level 1 down to the current lowest level only , those for deeper levels are hidden (grayed out). This is because calculating the number of nodes that would match the query at deeper levels is computationally intensive and would slow down the dynamic queries on nodes at the current level. At any time users can select the button to hide the grayed out nodes.
To summarize, the PDQ Tree-browser is a visualization tool for
hierarchical data that makes use of dynamic queries and pruning.
The PDQ Tree-browser uses two coordinated or tightly-coupled views
of the same tree, one a detailed view and the other an overview.
The user can select up to 3 attributes (numerical or textual)
for dynamically querying nodes at each level in the hierarchy.
Subtrees of nodes that do not match the query at their own level
are pruned out of the visualization. Thus, one can reduce a large
data set (with thousands of nodes) to a much smaller set, from
which good selections can be made.
4.3 Example Application: The UniversityFinder
We applied the PDQ Tree-browser to 2 applications, Network Management and the UniversityFinder. The latter is described here. The database organizes universities hierarchically; we have regions of the world, followed by states, then universities, and finally, departments. The UniversityFinder demo is available on videotape (Kumar, 1995).
Let us say that I am a high-school senior looking for universities that best match my needs. The PDQ Tree-browser initially shows all the regions in the world. I ask to see the entire tree to get an idea of the size of the database. The feedback indicators show that there are 740 departments in 286 universities in 62 states in 5 regions of the world (Figure 4).
I realize that panning this entire tree is not an easy task and return to the region level. Since I am only interested in regions where English is the primary language, I create a textual menu of primary languages by dragging-and-dropping the attribute Primary Language on to one of the empty slots under the Region level button. I select English from the menu and the South America turns gray, while USA, Canada, Africa and Europe remain yellow.
I now proceed to the state level. South America's subtree is not expanded. That subtree was pruned out since South America did not satisfy our criterion for primary language. However, South America is still shown in gray in order to provide context feedback.
I would really like to study in a state that is relatively safe, so I choose to query on the level of violence. I manipulate the range-slider to select only states with a violence index less than 65, leaving 37 states. I further reduce the number of interesting states by choosing only those with good traffic conditions. The number of states has now dropped to only 7 (Figure 5).
Now, I can look at the universities, but first I recapture space occupied by uninteresting states by clicking on the Hide Gray Leaves button. This gives me a more compact view that often is visible on a single screen. When I now go to the university level, only the remaining 7 states are expanded to show 25 universities.
I now reduce the number of universities by first eliminating those with high tuition and then setting the average SAT scores to closely match mine. I hide gray leaves once again and see that I am down to 7 universities: a couple in Arkansas, a couple in Ontario, etc.
Satisfied with this set of universities, I proceed to the department
level to closely look at departments (Figure 6) that best match
my interests and needs in terms of availability of financial aid,
4.4 PDQ Tree-browser Limitations
The UniversityFinder scenario demonstrates how the PDQ Tree-browser can be applied to everyday applications, in addition to complex applications like network management. The tool itself is general and can be used for any hierarchical data set with attributes for nodes at different levels of the hierarchy. However, the current implementation does have the following limitations:
4.5 Possible design alternatives
There are some interesting design alternatives that deserve special mention here:
4.6 Generalizing the design
This section attempts to generalize the PDQ Tree-browser design
so as to overcome some of its limitations and make it more generally
applicable to trees of varying structures and sizes. Ideas on
extending the query interface to allow specification of complex
boolean queries are discussed. The PDQ Tree-browser illustrated
the advantages of using dynamic queries and pruning while visualizing
trees as 2D node-link diagrams. The same advantages are there
to be had by other tree visualizations as well, e.g. treemaps
and Cone Trees. We illustrate this with examples for the treemap
4.6.1 Coping with varying structure and growing size: In
this section, we examine some of the issues that need to be addressed
in order for our design to be extensible to trees of varying structure
Performance issues: Our node-link layout algorithm, in
the worst case (i.e. when all nodes are to be shown), requires
2 complete traversals of the tree, Increasing levels displayed
requires more than 3 traversals of the entire tree, in the worst
case. Dynamically querying nodes at any level is more efficient
in that is only requires traversal of each node at that level
and not of the entire tree. Therefore, the system is likely to
slow down appreciably as the size of the tree increases, especially
for operations that require re-computation of the structure and
More sophisticated data structures and algorithms would be necessary
in order to minimize the performance deteriorations. For example,
one might only traverse the nodes visible in the detailed view
to evaluate dynamic queries. But this would mean that the overview
would not be tightly-coupled with the detailed view. Algorithms
for pruning could be improved so that only subtrees that have
been dynamically queried since the last structure change need
to be traversed. There are several interesting challenges that
remain to be solved, with respect to performance issues.
Interface issues: Ideally, the overview would always show
the entire data display in miniature, even if it means that no
details are visible; and the overview provides only global context.
However, when the size of the tree gets huge (say 50,000 nodes),
there is no way that one overview would suffice. Plaisant, Carr,
& Hasegawa (1992) found that in some cases, it might even
be useful to provide an intermediate view in addition to the overview
and the detailed view, when the detail-to-overview zoom ratio
is above 20.
Another feature that might be useful is to allow users to restrict
the nodes to be displayed before displaying them. For example,
if the user requests to see a new level with 20,000 nodes, the
system should present the user with the option of restricting
this set (feedback might be provided by displaying the number
of matching nodes) before displaying it.
4.6.2 Specification of general boolean queries: The current PDQ Tree-browser implementation allows users to specify an AND of queries on 0-3 attributes at each level. An improved interface would allow any number of ANDs, ORs and NOTs. But specifying complex boolean queries graphically is a challenge, as it may be difficult to interpret a set of graphical widgets, boolean operators and parentheses, even for experienced users.
The PDQ Tree-browser could accommodate complex boolean queries
by using the Filter-Flow metaphor in which queries are shown as
water streams flowing through sequential filters for ANDs, parallel
filters for ORs, and inverted filters for NOTs (Young & Shneiderman,
1993). The decrease in the breadth of the water stream while passing
through a filter represents the reduction in the data set due
to the corresponding query.
4.6.3 Pruning applied to treemaps: It was in fact, a treemap of a network hub that first highlighted the need for some mechanism to hide subsets of nodes / prune subtrees that were not interesting and recapture screen space. Uninteresting leaf nodes had taken up about 40 % of the total display space.
Figure 8 shows how pruning can help treemaps recapture screen
space allocated to uninteresting subtrees.
4.6.4 Dynamic queries and pruning applied to cone trees: Robertson
et al. (1991) describe ``gardening operations'' where the
user can manually prune and grow the view of the tree. Prune and
grow operations are done either by menu or by gestures directed
at a node. We believe that allowing users to make attributes-based
subset specifications in addition to these manual subset specifications
will help make Cone Trees more powerful and useful.
5 Usability Testing
In order to assess and improve the PDQ Tree-browser, usability
evaluations were performed, in which 8 subjects were given training
and asked to perform specified tasks using the PDQ Tree-browser.
The UniversityFinder data was used in the testing.
With each subject, the following procedures were followed:
Pilot testing was done with 2 graduate students, to try out the
tasks and get preliminary reactions. The tasks were refined, a
subjective evaluation questionnaire was designed, and the software
was tested with 6 subjects. The subjects were familiar with tree
structures and pruning, database querying, GUIs and direct manipulation
widgets. The subjects ranged from graduate students and undergraduate
seniors in Computer Science / Electrical Engineering to a faculty
The range of (7) tasks included:
Feature-based testing: Testing the specific features of the PDQ Tree-browser.
Task-based testing: Testing queries of varying complexity and type. This classification of queries is general and independent of this particular PDQ Tree-browser design and implementation.
The subjective evaluations revealed which features were liked
and which needed improvement:
5.4.1 Most liked features: When asked what was the one thing they liked the most about the PDQ Tree-browser, subjects responses included:
5.4.2 Subjective ratings: Users rated 22 aspects of the
PDQ Tree-browser on scales of 1 to 9, where 1 was the worst rating
and 9 the best rating. They also rated 3 possible features. For
example, users gave a mean rating of 8.7 (standard deviation =
0.5) to the usefulness of pruning, a mean rating of 8.2 (standard
deviation = 1.0) to the usefulness of dynamic queries, while the
ease of range-slider manipulation got a mean rating of only 4.8
(standard deviation = 2.4).
5.4.3 Suggestions for improvement
Subjects were also asked to make suggestions for improvement and
possible new features. All the suggestions were tabulated, resulting
in 19 possible enhancements to the system. Some of these are obvious
ones that would be inexpensive to incorporate into the system,
while others would require more effort:
(1) Ability to manually add or remove 1 or more nodes to the selection
subset resulting from the dynamic queries (by clicking, marquee
etc.). Thus, the ideal subset specification mechanism would allow
both query-based specification as well as manual specification.
Some issues relating to keeping track of these nodes and avoiding
complications later in the query process need to be addressed.
(2) Pruning and the Hide Gray Leaves feature were clearly well
liked by the subjects. Most of them used the latter frequently
to recapture screen space from uninteresting gray nodes and focus
the search. Removing nodes like this also made the system faster.
Many subjects agreed that a Hide All Gray Nodes feature would
be very useful. Some subjects also felt that the default behavior
could be to hide gray nodes (when levels changed) and users could
ask to see gray nodes (for context) if they were interested. This
suggestion should help make the views even more compact. But the
preferred default behavior depends on the structure of the tree
(fan-out at each level) and also the type of tasks; therefore
this requires more investigation. In fact, one subject liked seeing
the gray nodes at higher levels and used the Hide Gray Leaves
feature only once. He said the presence of the gray nodes at higher
levels helped in visually separating disjoint groups of nodes.
(3) Subjects found the response time of some PDQ Tree-browser
features to be slow, especially increasing levels when the number
of nodes was high.
(4) It was observed that subjects get somewhat disoriented when
the level of the tree was changed. This is because the layout
algorithm generates a fresh layout whenever the tree structure
changes, i.e. whenever more or less levels are requested to be
seen. It is felt that this problem can be significantly alleviated
by retaining the same current focus. For example, if the user
asks to see the University level while the state Florida is near
the center, the new view should be initialized to show universities
(5) Panning the detailed view by dragging the field-of-view in
the overview was found useful, but some improvements to the design
are required. One subject emphasized the need to always fit the
overview into one screen only, so that no scrolling of the overview
is required. As mentioned before, this is what we had designed,
but it wasn't enforced in this implementation. Another subject
suggested that users should be able to click anywhere in the overview
and have the field-of-view jump to that position. This would enable
fast coarse navigation. Fine-tuning could then be accomplished
by dragging the field-of-view.
(6) The catchiest quote received from one of the subjects was
this: "Drag-and-drop becomes a drag for experienced users,
so drop it!''. Some other subjects also echoed the feeling that
it might be easier and faster to just replace each drop area with
a menu of attributes at that level. Some other subjects enjoyed
the drag-and-drop mechanism to create and modify query panels.
(7) Ability to specify complex boolean queries involving ORs and
NOTs, in addition to the ANDs allowed currently.
(8) One task (Task 6) required subjects to look for departments
within 3 states, Wyoming, Wisconsin and Washington. This is basically
restricting states to Wyoming OR Wisconsin OR Washington. The
menu allows users to select 1-of-n textual values. A widget to
select m-of-n textual values needs to be designed and implemented.
(9) One subject wished that there was a way to search upwards
in the tree, i.e. to be able to query on universities and then
see the states which contained the selected universities. Another
subject made an interesting suggestion, that the PDQ Tree-browser
should allow users to hide certain levels on demand. For example,
if users are interested in looking at all universities in USA,
and do not care about the states they are in, it should be possible
to remove the State level totally and then get it back when desired.
6 A controlled experiment
With our guidance, three students, Robert Ross, Zhijun Zhang and
Eun-Mi Choi, conducted a controlled experiment to compare 3 behavior
alternatives for the PDQ Tree-browser (Ross, Zhang, & Choi,
1994). The UniversityFinder database was used in this experiment.
These behavior alternatives related to pruning subtrees of nodes
that did not match the query at their own levels. The 3 treatments
were as follows: (two of these were discussed in Section 4.5)
Full-tree: Subtrees of unselected nodes are shown in their entirety, but are colored gray.
Partially-pruned: Subtrees of unselected nodes are pruned out but the nodes themselves are shown in gray (This is the behavior option that the PDQ Tree-browser currently uses).
Fully-pruned: Subtrees of unselected nodes and the unselected
nodes themselves are pruned out.
It was hypothesized that subjects using the full-tree interface would take longer than those using the partially or fully-pruned interfaces. Although the full-tree allows for a static display of the tree, it's inclusion of all irrelevant nodes would probably slow down the response time, add too much useless information to the user's visual field, and create a need for excessive scrolling/panning of the displays to get to relevant nodes. Due to the longer task completion times and more difficult / complex searching required, users would also have lower subjective ratings for the full-tree interface than for the pruned interfaces. Further, the completion times and subjective ratings for the pruned interfaces would be approximately equal (differences would not be significant). It was postulated that the fully-pruned interface promised the most compact views, while the partially-pruned interface offered additional context feedback that might help in tasks requiring several iterations (to refine the query).
24 subjects were randomly assigned to use one (and only one) of
these three treatments (between-groups experimental design). They
performed a set of seven tasks. Afterwards, they filled out an
electronic questionnaire (QUIS) to subjectively evaluate the treatment
that they had used. The results (Tables 1 and 2) showed that the
times to complete the set of seven tasks were significantly different
at the 0.05 level with interface 1 being slower than 2, which
was slower than 3. In terms of subjective satisfaction, 2 rated
higher than 3, which was preferred to interface 1.
Table 1: Mean task completion times (in seconds) with standard
deviations in parentheses
Table 2: Mean subjective satisfaction ratings (normalized to 0
- 1, 1 being best) with standard deviations in parentheses
For total task completion time, an ANOVA gave F(2, 21) = 25.2 which was significant at the 0.01 level. Then, using pairwise t-tests, it was found that all results were significant at the 0.05 level. Therefore, in this controlled experiment, the fully-pruned interface was the fastest, while the full-tree interface was the slowest. The differences between the full-tree and the pruned interfaces were also significant at the 0.01 level, but the differences between the partially and fully-pruned interfaces were not. For subjective ratings also, the ANOVA produced an F that was significant at the 0.01 level. Also, all of the pairwise t-tests were significant at the 0.05 level. Users liked the partially-pruned interface the best and disliked the full-tree interface the most. Differences between the full-tree and the two pruned interfaces were also significant at the 0.01 level although differences between the partially and fully-pruned interfaces were not. In addition to the overall times for each interface, the times to complete each task were also compared.
The poor performance times of subjects using the full-tree interface might be partly due to poor system response times. Specifically, that interface had consistently (noticeably) slower response times, especially when expanding to the lower levels ("University'' and "Department'' levels).
The differences between the two pruned interfaces were not as clear cut; task performance was faster with the fully-pruned interface but the partially-pruned interface got higher subjective ratings. The fan-out of the particular tree used probably increased the differences between these two interfaces. Specifically, the UniversityFinder tree has a large fan-out (50 states in USA) from the Region level to the State level, and this might help explain why the fully-pruned interface had faster performance times than the partially-pruned interface (a significant proportion of states were gray and thus took up a significant proportion of screen space). It would be interesting to repeat the experiment with trees of varying fan-out.
This controlled experiment clearly showed the advantages of pruning.
The choice of whether nodes not matching the query should be grayed
or removed out is best provided via the interface to the users.
The PDQ Tree-browser visualization tool presents trees in two tightly-coupled views, one a detailed view and the other an overview. Users can use dynamic queries to filter nodes at each level of the tree. The dynamic query panels are user-customizable. Subtrees of unselected nodes are pruned out, a feature that usability testing and a controlled experiment, showed to be very useful. Possible enhancements to the PDQ Tree-browser were identified.
The concepts of dynamic querying and pruning are general enough
that they can be applied effectively to other existing tree visualizations
like treemaps and Cone Trees. These concepts are also extensible
to graph structures, but that would require careful thinking and
7.1 Future Directions
(1) PDQ Tree-browser refinements:
(2) General GUI extensions: Newer widgets need to be developed to enable users to specify multiple selections on textual attributes easily. Another useful widget would be an extension of the 2-box slider to a n-box slider.
(3) Extension to Graph Structures: The PDQ Tree-browser
extended dynamic queries to one class of non-flat data sets, i.e.
hierarchical data sets. The most general form of any data set,
is the arbitrary graph. Graphs have nodes connected via an arbitrary
number of links of different types. Each link represents a relationship
between (among) the connected nodes. We believe that PDQ Tree-browser
concepts of dynamic querying and pruning / selective growing can
be extended to graph structures as well. Instead of distinct levels
in the hierarchy, we would then think of distinct classes of nodes
(and even links).
7.2 Other tree-browsing research issues
(1) Semantics-based browsing: Generic 2D browsers (Plaisant, et al., 1995) treat the information space being browsed as images only. We believe that browsing of trees can be facilitated by taking advantage of the underlying structure of the tree. Fast navigation between siblings, up to parents and grand-parents etc., without having to manually scroll and pan would be useful options. Traversal of the tree in preorder, postorder and inorder, and tours of nodes marked either manually or by a query are interesting topics for investigation.
(2) Layout issues: There is a clear need for layout guidelines
for tightly-coupled overviews and detailed views, as the size
and fan-out of trees vary. For example, if the tree is wider than
deep (as was the case in the UniversityFinder scenario) then it
makes sense to have the tree drawn from left-to-right (or right-to-left)
and the overview to the left (or right) of the detailed view.
On the other hand, if the tree is deeper than it is wide, then
it might be better to have the tree drawn from top to bottom and
the overview below the detailed view. If the tree structure changes
frequently and has to be redrawn, then it might be a good strategy
to utilize the overview optimally by making the remaining tree
occupy the entire overview space. But this might lead to some
disorientation, as the zoom ratio will keep varying.
We are thankful to Dr. Michael Ball for his insightful suggestions, to Robert Ross, Zhijun Zhang and Eun-Mi Choi for designing and running the controlled experiment that was described in this paper, and to all the usability testing subjects who provided us with invaluable feedback. Thanks also to Edward Johnson for suggesting the name "PDQ Trees" for our system.
This project was supported in part by Hughes Network Systems,
Maryland Industrial Partnerships (MIPS), the Center for Satellite
and Hybrid Communication Networks, and the National Science Foundation
grants NSFD CD 8803012 and NSF-EEC 94-02384.
AHLBERG, C., SHNEIDERMAN, B. (1994). Visual Information Seeking: Tight Coupling of Dynamic Queries with Starfield Displays. Proceedings of the ACM Conference on Human Factors in Computing Systems, pp. 313-317, ACM., New York.
BATTISTA, G. D., EADES, P., TAMASSIA, R., TOLLIS, I.G. (1989). Algorithms for drawing graphs: An annotated bibliography. Unpublished Technical Report, Brown University, Computer Science Department, October.
BEARD, D. V. , WALKER II, J. Q.(1990 ). Navigational Techniques to Improve the Display of Large Two-Dimensional Spaces. Behavior & Information Technology. 9 (6), 451-466.
BUJA, A., MCDONALD, J. A., MICHALAK, J., STUETZLE, W. (1991). Interactive Data Visualization using Focusing and Linking. Proceedings of the IEEE Visualization ë91 Conference, pp. 156-163, San Diego, CA, IEEE Computer Society Press, October .
CARR, D., JOG, N. K., KUMAR, H., TEITTINEN, M., AHLBERG, C. (1994). Using interaction object graphs to specify and develop graphical widgets. Technical report CS-TR-3344, University of Maryland, Department of Computer Science, September.
CHIGNELL, M. H. , ZUBEREC, S., POBLETE, F. (1993). An Exploration in the Design Space of Three Dimensional Hierarchies. Proceedings of the Human Factors Society,, Santa Monica, CA.
GEDYE, D. (1988). Browsing the tangled web. Master's thesis report, University of California at Berkeley, Division of Computer Science, May.
HENRY, T. (1992 ). Interactive Graph Layout: the Exploration of Large Graphs. Ph.D. Thesis,, University of Arizona, Tucson, AZ, May.
HENRY, T. R., HUDSON, S. E. (1991). Interactive Graph Layout. In Proceedings of the ACM SIGGRAPH Symposium on User Interface Software and Technology, pp. 55-64, ACM, New York, November.
HOLLANDS, J. G., CAREY, T. T., MATTHEWS, M. L., MCCANN, C. A. (1989). Presenting a Graphical Network: A Comparison of Performance Using Fisheye and Scrolling Views. In G. Salvendy and M.J.Smith, Eds., Designing and Using Human-Computer Interfaces and Knowledge Based Systems, pp. 313-320, Amsterdam:: Elsevier Science.
KUMAR, H. (1995). Visualizing Hierarchical Data with Dynamic Queries and Pruning - The Tree-browser. In Catherine Plaisant, Ed., HCIL Open House '95 Video, Human-Computer Interaction Laboratory, University of Maryland, June.
KUMAR, H., PLAISANT, C., TEITTINEN, M., SHNEIDERMAN, B. (1994) Visual Information Management for Network Configuration. Technical Report CS-TR-3288, University of Maryland, Department of Computer Science, June.
MARCHIONINI, G. (1995). Information Seeking in Electronic Environments, Cambridge University Press, UK.
PLAISANT, C., CARR, D., HASEGAWA, H. (1992). When an intermediate view matters - A 2D-browser experiment. Technical Report CS-TR-2980, University of Maryland Department of Computer Science, October.
PLAISANT, C., CARR, D., SHNEIDERMAN, B. (1995). Image Browser Taxonomy and Guidelines for Designers, IEEE Software 12, (2), 21-32, March.
ROBERTSON, G. G., MACKINLAY, J. D., CARD, S. K. (1991). Cone Trees: Animated 3D Visualizations of Hierarchical Information. Proceedings of the ACM Conference on Human Factors in Computing Systems, pp. 189-194, New York:ACM.
ROSS, R., ZHANG, Z., CHOI, E. (1994). Using the Tree-browser to visualize information for database queries. Unpublished report, University of Maryland, Department of Computer Science, December.
SCHAFFER, D., ZUO, Z., BARTRUM, L., DILL, J., DUBS, S., GREENBERG, S., ROSEMAN, M. (1996). Comparing Fisheye and Full-Zoom Techniques for Navigation of Hierarchically Clustered Networks. ACM Transactions on Information Systems 14, (2 ).
SHNEIDERMAN, B. (1992). Tree Visualization with Treemaps: 2-d Space-Filling Approach., ACM Transactions on Graphics, 11, (1), pp. 92-99, January.
SHNEIDERMAN, B. (1994). Dynamic queries for visual information seeking. IEEE Software, 11, (6), 70-77.
YOUNG, D., SHNEIDERMAN, B. (1993). A graphical filter/flow representation
of boolean queries: A prototype implementation and evaluation.
Journal of the American Society for Information Science,
44, (6), pp. 327-339, February.
Harsha Prem Kumar received a B.Tech. in Textile Engineering
from the Indian Institute of Technology, New Delhi, India in 1992,
and an M.S. in Systems Engineering from the University of Maryland
at College Park, in 1994. He is currently a member of the Usability
Engineering group at Bellcore, Inc., Piscataway, New Jersey. His
research interests include information visualization, user interfaces
for network management, and interactive information systems.
Catherine Plaisant is assistant research scientist at the
Human-Computer Interaction Laboratory of the Center for Automation
Research at the University of Maryland. Her work centers primarily
on graphical user interface design and evaluation. She has worked
on various research projects related to hypertext, touchscreens,
public access information systems, information visualization and
many other domains. Plaisant received a general engineering degree
from l'Ecole Nationale Supérieure des Arts et Métiers
in Paris France, and a Docteur Ingénieur degree from the
Université Pierre et Marie Curie in Paris.
Ben Shneiderman is a professor of computer science, head
of the Human-Computer Interaction Laboratory and a member of the
Institute of Systems Research, all at the University of Maryland
at College Park. He organizes an annual satellite television presentation,
"User Interface Strategies," is the author of Software
Psychology: Human Factors in Computer and Information Systems
(Little Brown, 1980) and Designing the User Interface:
Strategies for Effective Human-Computer Interaction: Third Edition
(Addison-Wesley, 1997) and the editor of Sparks of Innovation
in Human-Computer Interaction (Ablex, 1993). Shneiderman received
his Ph.D in computer Science from the University of New York at