Currently, many people use spreadsheet programs at work and at home. Sometimes the spreadsheet data represents hierarchical (or tree) information. When data sets become too large to effectively extract information such as trends and correlations from them, then suitable visualization techniques must be found.

Fisheye Tables, Table Lens, and Treemap are similar in that they display an entire dataset to the user in one screen and that the data they use is tabular (in the case of Treemap, values from some of the table columns indicate each item's position in the hierarchy).

The Fisheye Table application uses the Fisheye data visualization technique to view tabular data. Data fields are organized into columns, where the user can sort by a single data field at a time, with the click of a mouse. To view the data on one screen, the Fisheye technique keeps a portion of rows in focus, while keeping the edges of the dataset in view, but out of focus to the user. Using the minimal movement of a mouse, the user can scroll up and down the rows to bring the edges in focus, and can easily scan the entire dataset. Each row represents a leaf node of the data set hierarchy. data that Treemap maintains. A snapshot of Fisheye Table is below.

Eureka embodies the Table Lens data visualization technique to view tabular data. Like the Fisheye technique, the Table Lens technique arranges data in tabular form, where columns can be sorted with a single mouse click. However, Table Lens initially gives the user a view of the entire dataset in compressed form, where individual data values are not viewed and scrolling is not necessary. Eureka accomplishes this by using colors to discriminate non-numeric data fields, and using variable length bars to discriminate numeric data fields. The resulting display allows the user to recognize patterns and trends among the fields instantly. To isolate specific data, Eureka also allows the user to filter the dataset based on any column's data values, and also uncompress and focus the rows that are of interest. Users have the option of anchoring columns, thus allowing them to sort other columns based on anchored columns' positions. A snapshot of Eureka is below.

The program Treemap embodies the Treemap data visualization technique to view hierarchical data. This technique visually maintains the hierarchical view of the data. Since the entire data set is viewed, the user has a powerful, fast, and efficient way to navigate through the data set, as well as the ability to recognize patterns and trends that may exist among the hierarchical levels. Treemap also utilizes filtering, sizing, and coloring features to allow the user to discriminate and isolate specific data easily. One popular commercial application of the Treemap technique is SmartMoney.com's Map of the Market. A snapshot of Treemap is below.

Researchers have considered the problem of visualizing tabular data sets that have more data entries than can fit on the screen using a spreadsheet program. One solution is fisheye views, which give the user the ability to focus on some rows of the table while preserving the context of the entire table by presenting the unfocused rows in smaller scale [1]. One program that uses this approach is Table Lens from the Xerox Palo Alto Research Center [5]. Another such program is Fisheye Table from the University of Maryland Human Computer Interaction Laboratory (HCIL). Fisheye programs allow users to investigate large tabular data sets more effectively than they would with a spreadsheet program [5, 10]. The Table Lens graphically differs from the Fisheye Table in that the Table Lens displays all the unfocused columns at the same compressed height whereas the heights of unfocused rows in the Fisheye Table taper off in proportion to their distances from the focused rows.

If the tabular data represents a hierarchy, then another solution is to display the tabular data as a hierarchy of nested rectangles in which clicking each leaf node reveals the attributes that the hierarchy does not convey; this is the approach of the Treemap from the University of Maryland HCIL [7]. The Treemap can show entire large data sets in one screen, instead of requiring a fisheye focus+context compromise. The Treemap program also allows users to encode attribute values using rectangle sizes and colors.

Another consideration in visualizing large tabular data sets is how user interaction can maximize the amount of relevant information on the screen. One interactive function is sorting. Sorting allows the user to cluster entries that are related by a given attribute so that it is easier to compare their values for other attributes. Another interactive function is filtering. Filtering lets users eliminate from view the entries that are irrelevant to the task at hand, leading to faster task completion [5, 6]. More rapid, incremental, and reversible interactive functions promote better user performance and satisfaction [11]. The Table Lens filters are inadequate according to that rule; they are slow and difficult to reverse, because the user has to open a dialog box and click a command button to apply each filter, resulting in a new window containing the filtered data. In contrast, Treemap filters are rapid and reversible, because the display updates for every incremental filter change that the user applies via on-screen range scroll bars.

Many factors influence how quickly individuals can interpret data presented visually. Individual differences such as experience, intelligence, physical ability play a role. Environmental conditions such as display quality, lighting, ergonomics, distractions impact cognition as well. Holding those and other related factors constant, what remains is how well a program's features promote data exploration how well a program's data representations aid cognition. The designers of the three products examined in this experiment have aided data conceptualization in many ways. This section focuses on aids of perceptual redundancy, focus+context, and dynamic queries.

Sebrechts describes perceptual redundancy as "two different dimensions,...both represent[ing]...a common property" [12]. In Treemap, perceptual redundancy occurs when users set the size by and color by legends to the same property in a query. The sought after result is perceived with greater ease as both the largest and most brightly colored rectangle on the screen. This redundancy makes it possible for a user to finger an item faster than one would have had using only a single encoding. Redundancy is provided in Table Lens by allowing the user to both view the relative lengths of bars in a chart and sort them as well. The result allows one to more quickly perceive maxima and minima because of both their relative lengths in a column and their relative positions at the extremes of a column. Similarly, in Fisheye Tables, one can sort a column so that one can more quickly perceive maxima and minima based on its position at the extremes of a column.

Treemap and Table Lens present data graphically so that the user can grasp at a glance quantitative elements and relationships. For example, a user may identify correlations from graphical representations of datasets. In Treemaps, if one were to color nodes by flow and size nodes by severity of problems it may become visually apparent to a user that output value drops below a certain amount only when the number of problems rise above some value. In Table Lens and Fisheye Tables, users can identify "cross variable correlation" when they sort a column and thereby the values of another column become mostly sorted as well. In addition, statistical distributions become apparent if one knows how to recognize shape and skew [14].

An outstanding feature for facilitating cognition in Table Lens and Fisheye Table is focus+context, i.e. the ability to focus on more pertinent information while keeping less pertinent information on the screen solely to provide context. In Treemaps, cognition is aided with dynamic queries, the ability to rapidly, reversibly, and incrementally apply and view the results of filtering operations. One benefit of these features is that the user can adjust the breadth and depth of data representation to aid comprehension and exploration. Given that "cognitive research has shown us that people have trouble comprehending even simple displays and understanding relatively simple data" it stands to reason that dynamic queries and focus+context, which simplify complex displays will reduce users' cognitive load and promote understanding [13].

In Treemap, users can directly manipulate and simplify the display without losing context by filtering with double sliders. Such dynamic queries focus users' attention on the results by graying out nodes that fail any filter. Treemap preserves context by displaying all nodes in the data set. Users tend to maintain their orientation with respect to the data and the program's interface when context is preserved.

This experiment compares the effectiveness and satisfaction of people using several different techniques for visualizing 100 units that form a shallow, fixed-depth tree; each unit possesses several numerical attributes. The experiment varies a single independent variable (visualization technique) with three treatments: Fisheye Table, Table Lens, and Treemap. Each subject answered the same four questions about the data using each treatment. Below is a diagram illustrating part of a typical hierarchy of the data. The dependent variables are time to correctly answer of the question, the number of incorrect answers given before the correct answer, and subjective satisfaction.

Previous studies suggest the following predictions:

1. Sorted table column v. size- or color-coded 2-D space
   For data sets containing about 100 units, usually subjects can sort a table by a column then search the column for a specified value faster than they can scan an unsorted 2D space for a unit possessing that value, even if the value is size- or color- encoded. (However, scanning the 2D size- or color-encoded space might be faster for orders-of-magnitude larger data sets, especially if the differences in size and color are large enough.)
   
   

2. Filtered v. unfiltered
   Given a sorted table column, subjects take less time to filter out irrelevant rows (using a program feature) then find a specified value by visually searching the remaining rows than they take to find a specified value by visually searching the rows without filtering them first.

A prediction not based on previous research is that, given a sorted table column, subjects take less time to find a specific value in the column using continuous fisheye techniques (such as the Fisheye Table) than they do using discrete fisheye techniques (such as the Table Lens), because, with continuous fisheye techniques, users can see more data in nearby unfocused rows.

Combining these predictions leads to our hypotheses. In the hypotheses, the concept of better performance means less time to correct completion and less or equal number of incorrect answers. Each question tested a different aspect of the treatments in an attempt to clarify the strengths and weaknesses of each treatment. The hypotheses are:

1. On question 1, which requires determining which unit possesses the maximum value for an attribute, subjects will perform best using the Table Lens and the Fisheye Table techniques (because they let the user sort by column). Users will not perform as well using the Treemap technique (because its color- and size-encoding will not help as much as sorting).
   
   

2. On question 2, which requires determining how many units meet range restrictions for multiple attributes, subjects will perform best using the Treemap technique (because of its rapid, incremental, and reversible filters). The next most efficient treatment will be the Table Lens technique (because it also has filters, but they are less rapid and less reversible). Subjects will perform least well using Fisheye Table technique (because it lacks filters).
   
   

3. On question 3, which requires determining the unit that possesses the minimum value for an attribute out of a set of units that meet a value range restriction for another attribute, subjects will perform best using the Table Lens technique (because it provides filtering and sorting). The next most efficient treatment will be the Treemap technique (because it provides filtering but not sorting). Subjects will perform least well using the Fisheye Table technique (because it has no filtering).
   
   

4. On question 4, which requires repeatedly finding units that possess specified values for an attribute, subjects will perform best using the Fisheye Table technique (because it provides sorting and continuous fisheye layouts). The next most efficient treatment will be the Table Lens technique (because it provides sorting and discrete fisheye layouts). Subjects will perform least well using the Treemap technique (which lacks sorting but provides a hierarchical view that might simplify this task).
   
   

5. Subjective ratings of satisfaction will rank the treatments in this order with statistically significant differences: Table Lens, Treemap, Fisheye (because that is the ranking of the treatments by decreasing number of features that the interface provides).
   
   

6. Subjective ratings of ease of use will not contain any statistically significant differences between the treatments since all the programs provide straightforward interfaces to the features they provide and to the styles of interaction that they promote.

Pilot Study Results

The objectives of the pilot study were to ensure that subjects could understand the training and could answer each question in a reasonable amount of time and also to determine how much time to allot for each subject to complete the experiment. Running the pilot tests also provided the experimenters with practice running the experiment. Three subjects participated in the pilot study. Originally, the pilot study included five questions. The most drastic pilot study result was that two of the questions were extremely difficult for subjects to complete in a reasonable amount of time. One of these questions required too much scrolling with two of the treatments. Restricting the number of units that this question involved made it manageable. Another question required users to sum values while scrolling, which proved too difficult a task for all the pilot subjects. The question, which required subjects to sum attribute values for groups of units, did not seem to be worth the frustration it caused the subjects, so it was eliminated from the experiment, leaving four questions. The pilot subjects commented on certain weaknesses in the training materials. Revisions of the experiment materials included suggestions from pilot subjects, such as explaining how to reset each visualization to its original settings in order to start each question with a clean slate. Other suggestions included spending more time on how each visualization presented the hierarchical aspects of the data. The maximum time that any pilot subject took for training was 15 minutes, and the maximum time that any pilot subject took for answering the questions for all three treatments was 15 minutes. Taking into account training and testing time and about 5 minutes for miscellaneous activities, such as filling out consent forms or questionnaires, gave a total experiment time of 35 minutes.

Subjects

Eighteen subjects took part in the experiment. All subjects had at least one year of experience using a window manager such as Microsoft Windows 95 or Apple MacOS.

Materials

Before performing any tasks, subjects filled out a three-question background questionnaire that asked for each subject's sex, color-blindness (if any), and years of experience with several similar window managers (such as Microsoft Windows 95 or Apple MacOS). The training materials for the experiment included an overview of the data (i.e., the structure of the hierarchy and the types of numerical attributes that each unit had), as well as an introduction to the features of each of the three visualization programs. Some of these program introductions demonstrated program features via short walkthroughs of how to solve simple questions using them. The four questions for the tasks were on a single sheet, and subjects could determine the answer to each question using any of the features of each visualization program that the training materials described. After completion of the tasks, each user filled out a brief subjective satisfaction form. The training questionnaires, materials, and questions are in the appendices.

Procedures

Each experimenter ran the experiment with six subjects, one with each permutation of the treatments.

Sample Solutions

Sample solutions to the questions, including screenshots at each stage are here.

Results

This experiment collected information on the time it took 18 subjects to complete 4 tasks, their error rates, and their responses to two subjective questions. The primary measure of performance was the time it took subjects to correctly solve questions using three different tools. Microsoft Excel was used to analyze the data. For each task and subjective question, a one-way ANOVA was generated. If the analysis indicated that within the three treatments there was a statistically significant result with respect to an alpha value of 0.05 then a paired t-test was employed. An analysis of the paired t-test indicated which pairs of treatments supported the statistically significant result. The appendices include the raw data.

Performance Time

Error Rate

Subjective Questions