Uncertainty Visualization Study Group Notes – 10/22/2012

Discussion

---

MacEachren Paper

• Arbitrary assignment of high uncertainty and low uncertainty

• The second experiment is subjective rating of undefined concept – improperly conflates intuitiveness (logical channel) and effectiveness
  • Uncertainty not tied to any context, dealt with only as an attribute of the data
    • In fact, the only attribute: uncertainty is meaningless when not referring to an attribute
  • Intuitiveness rating used to determine usefulness

• Some renderings of size are better than some renderings of color – can’t just compare the two without discussing different encodings of each…

• Better questions to answer:

  • “How does the visualization of uncertainty influence reasoning and decision-making in problems in which uncertainty is important?”
    • true context for uncertainty
  • “What symbolization methods work best when sign-vehicle represents both data and uncertainty?”
    • provides us a way to practically answer whether one method or another is more accurate

Uncertainty Framework and Preliminary Experiments

• In proposal, by fiat, we adopted a very particular way of specifying the kind of uncertainty we’re going to visualize – We are not dealing with sources independently, rather all sources of uncertainty are folded into a single probability distribution over outcome set.
  • [Miriah] We are ignoring the role of the sources / context in informing decisions – should we make more explicit?
  • Does the source of data and the source of uncertainty affect the effectiveness of particular representations?
• Assertion: 2 problems to solve – both must be explicitly specified
  1. What properties of pdf are we encoding? or What is the characterization of the pdf?
  2. How are we encoding those properties? or How are we encoding that characterization?
• Start with absolute simplest case, characterized by 1D random variable
  • Not b/c it’s interesting problem but b/c it is non-trivial and well formed
  • How do you know what constitutes a ‘right answer’? We’ve chosen our problem in a way that it is mathematically defined (e.g. maximum likelihood), rather than just subjectively defined
    • the task requires incorporation of uncertainty information in order to give the correct response
• [Ross] One-dimensional representations are very different from two-dimensional representations, both of which are very different from from three dimensional, etc. – The results of this study may not be very generalizable
  • The methodological framework for conducting the studies will still generalize to higher dimensions
  • This study will act as a basic threshold test of whether people can even cognitively process uncertainty information properly
  • We should include qualification language in our results - be honest about possible lack of generalizability

• [Miriah] We should think more about the variety of characterizations that we can produce from a given pdf

• What other sorts of task types (besides those already mentioned) can we produce to test uncertainty?

• Would this be a good Visualization paper? Would this be a good Cognitive Science paper?