Intelligent Enterprise

Better Insight for Business Decisions

RSS
Webcasts
Whitepapers
Subscribe
Home

June 26, 2000, Volume 3 - Number 10

The Beholder's Share

The Eye's Role in Visualization Technologies

Imagine the tremendous power of seeing pictures that isomorphically portray the essential relations of your business in ways your mind’s eye would construct them, images that you can rotate, move, and compare in ways your intuition naturally compels you. It’s a power that could lead to a constant stream of useful business ideas. Is this kind of visualization in data mining’s future? Or will we never do better than “lies, damn lies, statistics, charts, and graphs,” to paraphrase Mark Twain?

With increasingly more tools that let users produce custom images, we need to know how to choose our methods well. In other words, we must know how to translate correctly millions of records of business data into shapes and colors that fully unleash the brain’s sophisticated, multivariate pattern-recognition and manipulative talents.

I argue that, to find the right answers, we need to know about the very nature of seeing itself. There are two main parts to successful visualization methods:

1. When mapping from data to image, retaining the elements is essential to the problem. That is, you make sure the key business relations get appropriately translated into a visual metaphor — as in VisualMine by AI Software. Figure 1, shows VisualMine’s representation of different levels of cash flows between regions of Italy, transformed into a 3D spread of differently colored pipes that help identify extreme instances indicative of money laundering.

FIGURE 1 Flows of cash between regions, seen as pipelines.

2. Knowing specifically how the brain uses the given stimuli — how, for example, people interpret the streaks of colored paint on a flat canvas shown in Figure 2, as a warm summer day.

FIGURE 2 Why does the eye see a warm summer day from streaks of paint on a 2D canvas?

The former is a complex subject in its own right. Translation usually entails loss and distortion, as well as amplification. The key is to amplify and maintain what is essential to a given need. For example, how do you think the size of Africa and the United States compare? Most people do not know Africa has almost four times the land area of the United States. In mapping the surface of Earth onto a flat surface, it is possible to preserve only certain geometric relations. The commonly used Mercator projection, which preserves angles and is thus useful for navigation, greatly distorts sizes and is therefore not useful for understanding comparative land areas.

In this column I will focus on the second part: knowing how the brain responds to visual information.

The human eye can be remarkably powerful: It can tell when two end-to-end lines misalign by the minutest fraction. Or, it can totally blunder: The moon looks tiny when overhead, but immense when it’s a backdrop to a town on the horizon.

When are business decisions based on images similarly precise or illusionary?

The exercises, along with Figure 3, in the sidebar “Misperception” demonstrate some examples of pitfalls in visualization. The very mathematics of seeing require the viewer to make assumptions.

FIGURE 3 Some visualization methods provide illusory impressions about the business data.

The human mind first presumes simplicity; that samples tend to be typical and that there is continuity in processes, as you can see in the sidebar exercises A and B. When the world is not this simple, people often err. Take the most typical statistical blunders as examples: “I have just won four rolls of the dice. I must be lucky. Triple my bet.” “The fish (or the customers) are really biting at this location. Let’s stay here.” (But, if this sample is too small or atypical, you really know nothing about future success there.)

Furthermore, in a visualization, virtual reality’s clues may not be internally consistent — and even if they are, the eye can misinterpret them. For example, Julie Harris at The University of Newcastle conducted an experiment requiring subjects to catch a ball in a virtual reality environment in which parallax and size-change clues did not correspond to each other as in reality. She discovered that subjects decide when to close their hands on the incoming virtual ball based on whichever signal indicated the ball getting there first.

The literature on the heuristics and bias of human interpretation is extensive, and many areas of business and finance apply to them. For example, the sophisticated National Opinion Research Center in Chicago is highly aware of how presentation, such as how questions are posed, greatly affects results in surveys.

Consider, “How many countries, more or less than 27, are in Africa?” Whatever the answer, it will be biased toward 27. If you ask, “more or less than 16,” the answer will be biased toward 16.

Many comparable “anchoring biases” affect people’s sight: We see what we expect, not exactly what is there.

Even our perception that we see the world evenly laid out is an illusion. Julie Harris has shown that we overestimate the angle of objects coming toward us, often by a factor of 10. That is, subjects can believe a difference of only three degrees is 30 degrees!

Inxight’s Site Lens Studio, software that maps Web sites, museums, or even libraries (see Figure 4), mimics and takes advantage of this focus-expanding, periphery-summarizing tendency of the eye. When you click on any region of the map, it expands to give details while surrounding areas contract and offer only higher-level summary information. Check out www.inxight.com to give it a try.

FIGURE 4 A map that explodes details where you touch.

It is time for more makers and users of visualization approaches to become aware of the extensive literature on seeing, which not only comes out of brain and vision laboratory research but also from philosophers, artists, and art historians. (The latter two, being professional examiners of looking, often figure out properties of the eye long before the scientists.)

Is the 3D update of the highly multi-functional software DataScope by Mindmaker Inc. (see Figure 5) better than the 2D version? The 3D pie, though gloriously delicious, does not do what pie charts should do: namely, clearly show percent differences of the whole. Nor can the user perceive real distances and relations between points in the connected 3D scatter plots shown.

FIGURE 5 Is 3D easier to read than 2D?

TowerView by HighTower Inc. handles the ambiguity of size at different distances in perspective extremely well, by double coding it with color. (See Figure 6.)

FIGURE 6 Instantly seeing the mission-critical needle in the haystack.

One of the key elements of successful communication, particularly in ambiguous and noisy contexts, is redundancy. Subject-verb agreement is an example of signifier redundancy in natural language. Additionally, the human senses are highly tuned to noticing exceptions to the simple and expected, such as a shooting star in the night sky or the odd noise in the engine. TowerView takes advantage of this tendency in order to let users instantly notice critical ranges in a few variables out of thousands. This feature has been used to rapidly pull patients in danger during large clinical trials and by NASA to instantly find a failing satellite component among the thousands in orbit.

In Figure 7 you can see an example of the simple genius the Temple MVV software by Mihalisin Associates Inc., which uses multiple nesting of dimensions to display high-dimensional information (here, merely five dimensions) in easy-to-view 2D. You can derive a lot of information from the visual; for example, you can instantly see a glass ceiling for women’s salaries in contrast to men’s. Colors differentiate education ranges, with violet the highest and yellow the lowest. Men are on the left half, women on the right, with seven age ranges for each sex. The center of the bars indicates salary levels with the distance between the top and bottom of a bar the standard deviation.

FIGURE 7 The glass ceiling in women’s salaries — seeing five dimensions with two.

Of course people can perceive many dimensions of information instantly in even the slightest variations of, say, facial expression and body form. Considering how sensitive our visual sense is, the best products now available tap only a small fraction of the power of visual presentation. Perhaps someday, data mining will entail just sitting down to a full-length animated movie (like the ones by Pixar) and afterward knowing exactly what business move to make. But there I go, thinking in pictures again.

MISPERCEPTION

Some visualization techniques can mislead the user

Unless you take into account the substantial “beholder’s share,” as art historian E.H. Gombrich puts it, in image interpretation, a visualization method could easily be the source of highly deceptive conclusions.
These exercises help illustrate this point.
Take out your pencil. Look at Figure 3.
Questions:
A. Draw circles around the clusters. Where would you most likely expect new points to fall and not to fall?
B. Draw in the trends. Where would you expect the next values to be?
C. Which of the two small inner squares is lighter?
D. Is the horizontal or vertical line longer?
E. Do you see a square?
Answers:
A. The points are just randomly placed. Any clusters you notice are just arbitrary artifacts of this sample. Any new points in this process could go anywhere.
B. The process is just a random walk, neither straight nor wiggly lines have any value in forecasting where the next point will go.
C. The small squares are both the same color. Color perception, as with other forms of interpretation, is highly context-dependent. I have seen little concern for this issue in coloring map or image elements with numerically driven color scales. It’s unlikely viewers will assess the right value for each color, especially when any region might have any of a number of unpredictably colored neighbors.
D. The vertical and horizontal lines are the same length. Again, context is an important component of assessing size and distance. This assessment is especially difficult with pseudo-3D representations of perspective, in which the standard practice of mapping values onto the height of objects, such as bars, can create confusing pictures.
E. We see a square when there is none. Eighty percent of neural synapses entering the first vision areas come not from the eyes, but from higher-level cognitive areas of the brain. Expectation and context heavily bias what we perceive.