Perception

(Difference between revisions)
 Revision as of 03:26, 11 February 2011 (view source) (→Siamak Faridani - Feb 10, 2011 09:13:53 pm: new section)← Older edit Revision as of 03:35, 11 February 2011 (view source) (→Sally Ahn - Feb 10, 2011 09:34:36 pm: new section)Newer edit → Line 79: Line 79: One idea that I might for parallel coordinates was to place dimensions next to each other based on the correlation of data in each dimension with the other dimension. Or alternate between negative and positive correlation meaning that the two neighboring dimensions can have a highly negative correlation. Additionally I am still not convinced that connecting data in neighboring dimensions is a good idea. Lines in bump charts for example bring no valuable information to the visualization. And in parallel coordinates connecting links make the visualization crowded, ugly and hard to interpret. One idea that I might for parallel coordinates was to place dimensions next to each other based on the correlation of data in each dimension with the other dimension. Or alternate between negative and positive correlation meaning that the two neighboring dimensions can have a highly negative correlation. Additionally I am still not convinced that connecting data in neighboring dimensions is a good idea. Lines in bump charts for example bring no valuable information to the visualization. And in parallel coordinates connecting links make the visualization crowded, ugly and hard to interpret. + + == Sally Ahn - Feb 10, 2011 09:34:36 pm == + + Yup, the change blindness got me too; I actually never even got beyond the airplane one (until I read Karl's comment).  This and Simons and Rensink's findings that "changes…can…be missed, particularly when the changes are unexpected" places an additional burden on a good visualization, especially in exploratory data analysis, since changes and comparisons are what we are most interested while we iterate in search of meaningful relations hidden in the data.  I wonder if nonphotorealistic renderings of some of the change blindness examples would make the changes more noticeable.  Does the realism of photographs lead one to expect missing things to be there (in the case of the airplane example, the engine) and hence miss the change?  There are specific types of visualizations (such as maps…consider how much harder it is to spot places of interest in a satellite view) where NPR can clarify visualizations; I am not sure how we would apply this to data visualization design, when there is no such thing as "photo realistic" rendering. + + The authors also mention research on the effectiveness of brush stroke properties, but it is unclear to me how brush strokes, which combine color, orientation, width, length--so many features--in an undefined manner can be analyzed for effectiveness.  I am not very convinced by Figure 17; the shape, thickness, and length of the "brush strokes" strike me as the main influence; in fact, I may never have known that the visualization on the left were simulated brush strokes at a first glance.

Revision as of 03:35, 11 February 2011

Lecture on Feb 2, 2011

Contents

• Perception in visualization. Healey. (html)
• Graphical perception. Cleveland & McGill. (jstor)(Google Scholar)
• Chapter 3: Layering and Separation, In Envisioning Information. Tufte.

• Gestalt and composition. In Course #13, SIGGRAPH 2002. Durand. (1-up pdf) (6-up pdf)
• The psychophysics of sensory function. Stevens. (pdf)
• Crowdsourcing Graphical Perception: Using Mechanical Turk to Assess Visualization Design. Heer & Bostock. ACM CHI 2010. (html)

136.152.168.242 - Feb 09, 2011 01:56:43 pm

Regarding Figure 29 in the Cleveland & McGill paper, I would argue that this graph and a more traditional heat or patch map serve two different purposes. Although in this map it is easier to compare the murder rates from one state to another it does not take advantage of the preattentive processing described in the Healey paper. A traditional heat map gives a good overview of quantitative variables but it allows for only minimal comparison of similar values. This type of map gives a poor overview of quantitative variables and requires postattentive processing for finding anomolies, but allows for much more nuanced comparison of individual states' values even when the difference between them is minimal.

Michael Cohen - Feb 09, 2011 05:03:53 pm

I was frustrated by the "perceptually-motivated" visualizations featured in the Healey article, particularly the electoral map at the top and the weather history map (Fig. 17). I think this frustration spilled over into our discussion of distinguishing orientation and length in class and led me to question whether using both together would be a good idea in practice. I think Healey's visualizations are impressive in that they demonstrate that you can cram so many dimensions of data into a single visualization and have it all be fundamentally perceptible. However, in practice I think both visualizations are difficult to read and draw useful conclusions from. In the electoral map, the depth of each state corresponds roughly to its importance in the election, but since depths obscure other depths they are very difficult to compare except between geographically adjacent states. The depth dimension also causes some states to obscure others (e.g., California & Oregon). In the weather map, the use of size and density of strokes to encode some information means that the Southeast has significantly lower data density than the North on dimensions encoded by color and orientation. It's also difficult to disentangle density and size visually. I think the weather map, in particular, could have been better handled with small multiples.

To me, these visualizations illustrate that more information in a single display isn't always better, even if your perceptual theory argues that all of the information is available. The interference between dimensions is an crucial factor.

Saung Li - Feb 09, 2011 06:49:58 pm

I like how this topic of perception combines the fields of psychology and visualization. I did not know there are fields such as "psychophysics." The idea of preattentive processing, in which visual properties are detected quickly via a low-level system, can help direct us in our design of graphics. We are good at detecting differences in color and shape; after playing around with the Java applet it seems like differences in color are easiest to spot, though different shapes still work well. The change blindness phenomena shows us how little information we actually retain from looking at an image. I found it extremely hard to find the differences in the GIFs-- I could only find one if I focused on one small part of the image, and doing this can be quite painful to the eye. These ideas suggest that we should not clutter too much information into a graphic unless we make use of preattentive cues effectively, so that readers can pick up the most relevant information quickly.

Boheekim - Feb 09, 2011 10:12:51 pm

I just wanted to add a couple thoughts to today's class. I wanted go back to the trade off we face when we decide to make accurate visualizations or visualizations that are accurately perceived. Perhaps if we are only to absorb the data rather than analyze it, maybe it would be better to work towards making visualizations that are accurately perceived. Also on pie charts: I know we discussed why pie charts are often a poor data visualization. Do we use them because they are so familiar, because they are more aesthetically pleasing or for some other reason?

Lastly, the Healy paper talked a bit about directing the eye of the audience to the areas that are important. I think it's important to consider what mediums we are publishing the data visualizations in then, taking into account eye tracking when it is web for example.

Natalie Jones - Feb 10, 2011 01:51:39 am

I find the work on perception to be fascinating, and particularly the findings about misperceptions as they pertain to visualizations. It also makes me think about accuracy in visualizations in general. It seems important before starting a visualization to decide what accuracy means for that particular dataset and audience, and how adamantly it might be necessary to stay faithful to the data in order to get the intended point across. In some cases, it might be much more important than in other cases to be orthodox about scaling the data "correctly," and I could imagine that in some cases it might actually make sense to choose design over data, or at least meet in the middle somewhere. I also feel like I see a lot of visualizations that don't even seem to try hard at all to make certain aspects of the visual representation correspond to data, and I wonder if there are visualizations I've seen that appear attractive and interesting, but that may not have communicated as much as I thought they did, or as much as they could have. It's certainly something I'll be more on the lookout for now.

Karl He - Feb 10, 2011 02:40:31 am

I was particularly amused by the images on change blindness. It took me at least five minutes to realize that the airplane engine was missing in the second frame of the animation.

This phenomenon makes clear that human perception is extremely important for visualization. It doesn't matter how accurate your visualization is if it is interpreted incorrectly. Techniques such as apparent-size drawings I am hesitant to utilize since it involves distorting the data. However, it is a valid technique as long as nothing is being exactly measured. Apparent sizes of bars on a bar chart would likely be a bad idea, since people may try to measure exact sizes, but for the example given of dots on a map representing newspapers delivered, I feel it is valid since the disc sizes aren't meant to exactly correlate to the numbers anyway.

While it should be used with caution, it is a good idea to distort data for human perception to get the correct message across, especially if it is difficult to interpret without aid.

Dan - Feb 10, 2011 11:31:27 am

I found it interesting that in order to develop a scientific approach to visualziation, Cleveland and McGill use subjective methods involving human test subjects and attempts at characterizing graphical perception. I think it's hard to quantify a person's ability to decode graphical information mainly due to the background of the individual. If they are a scientist, they may parse more information. Or a graphics artist may think differently than a doctor. However, it could be assumed that all people have same graphical perception. Also, I think their method of comparing the theory to the experimentation results is something that justifies this. The perceptual buildings blocks that the authors coin “Elementary Perceptual Tasks” is also a great way to build a basis for understanding and quantifying the decoding of graphical data. I also found it interesting that subjects had a tendency to give answers that were in multiples of 5 when given questions with numerical answers.

The perception in visualization article was very engaging. I liked the first example of the 3D united states map with the various colors, textures, and depths to represent many variables. Also, the examples where you use comparison to see data stand out, for example, a red box in the middle of blue circles sticks out to the human visual system immediately, where if the box is blue it would not. The concept of preattentive processing can be used to create intuition for user interface design. This is relevant to my work as well as creating visualizations that are easy to decode for a viewer.

Matthew Can - Feb 10, 2011 04:35:31 pm

I enjoyed reading the paper by Cleveland and McGill. Not only do they have a scientific approach to building a foundation for graphical methods, but through their examples they also provide a lot of insights into how their theory can be applied to create better statistical graphics. Based on their ordering of the elementary perceptual tasks, they suggest redesigning the divided bar chart as a grouped dot chart. Likewise, "a pie chart can always be replaced by a bar chart." Humorously, they resolve the "Bar-Circle Debate" by stating that both are poor forms of graphical display that can be replaced by better ones.

What was more novel to me was their redesign of the shaded statistical map as a framed-rectangle chart. The authors provide a few reasons for this redesign. The most obvious is that framed rectangles are compared using position along nonaligned scales, which is higher on the ordering than shading comparison. A more subtle reason has to do with our ability to perceive clusters. The authors note that with the patch map (Figure 5), our ability to perceive clusters depends heavily on the shading scheme and that we are more sensitive to differences on one end of the scale than the other. This is related to the concept of Just Noticeable Difference (JND) that we talked about in class. Since JND is a function of ratio rather than magnitude, it means we have a tendency to notice clusters at the high-intensity end of the scale (that is where differences in ratio are smaller and so appear to "blend"). The higher accuracy framed-rectangle chart helps to combat this problem, though the authors suggest other numerical schemes that help even more.

On another note, we discussed in class the tendency for people to misperceive actual values of area and volume. In particular, we talked about the trade-off of representing data with actual area versus "perceptually correct" area. Cleveland and McGill argue against the latter for at least one reason, which is that the factor of misperception varies among people.

Sergeyk - Feb 10, 2011 07:30:40 pm

Like Michael Cohen, I too find the weather history map slightly offensive. I don't understand what comprehensible claim it is making, and thus I don't feel that it significantly lowers the entropy of the data. I could just look at the table which generated this visualization and be about as equally confounded. In fact, if I had a hunch that I wanted to confirm about this data, such as that it is both rainy and hot in Florida, I would probably prefer a well-designed, aggregated table to this visualization. And if the claim being made is some complicated, possible non-linear relationship between the three variables (that might further depend on the geography), then I would much rather comprehend this claim as a sentence or paragraph, and see confirmation of it in separately-plotted visualizations and in statistical analysis.

We can cram as many variables into a visualization as perceptually possible, and do it in the most accurately perceived way. But if the graphic does not come with a story, then it is mostly pointless. If it does come with a story, then we should be able to verbally express that story, and back it up with statistical analysis. The figure should not be the sole component of the communication, and consequently it does not have to be a perfect presentation of the data.

This is why I'm skeptical of trying to find the most perceptually correct ways to represent the most data--I would rather see presentations of data that speaks for itself loudly enough that a slight re-mapping does not really matter. If that's not true, then there's no reason I should be presented with the data.

But this is all about communicative uses of visualization. For other uses, for example dashboards, correctly perceiving the data is very important--but there should never be so much data to perceive so as to exceed our capacity of sumarizing it in a sentence or two.

Mhsueh - Feb 10, 2011 08:18:33 pm

I'll admit it. Change blindness totally got me. I showed my roommate, who has a pretty keen eye, and even he found it difficult to spot the changes that would otherwise seem obvious. All of the presented theories in the Healey reading that explain change blindness are convincing and very intriguing in my opinion; obviously there is still much we don't know.

Regarding the comments about cramming numerous variables into visualizations: I agree generally that the resulting visualizations are probably not the most effective. But I don't actually think Healey recommends this practice at all -- at least he never explicitly advocates it. It seems he does not present those images necessarily as examples of good visualizations, but rather of perceptually motivated visualizations, incorporating aspects of pre-attentive processing and feature hierarchy. It is similar to the way he shows example images related to the numerous pre-attentive processing theories while not actually endorsing any single one.

Siamak Faridani - Feb 10, 2011 09:13:53 pm

One idea that I might for parallel coordinates was to place dimensions next to each other based on the correlation of data in each dimension with the other dimension. Or alternate between negative and positive correlation meaning that the two neighboring dimensions can have a highly negative correlation. Additionally I am still not convinced that connecting data in neighboring dimensions is a good idea. Lines in bump charts for example bring no valuable information to the visualization. And in parallel coordinates connecting links make the visualization crowded, ugly and hard to interpret.

Sally Ahn - Feb 10, 2011 09:34:36 pm

Yup, the change blindness got me too; I actually never even got beyond the airplane one (until I read Karl's comment). This and Simons and Rensink's findings that "changes…can…be missed, particularly when the changes are unexpected" places an additional burden on a good visualization, especially in exploratory data analysis, since changes and comparisons are what we are most interested while we iterate in search of meaningful relations hidden in the data. I wonder if nonphotorealistic renderings of some of the change blindness examples would make the changes more noticeable. Does the realism of photographs lead one to expect missing things to be there (in the case of the airplane example, the engine) and hence miss the change? There are specific types of visualizations (such as maps…consider how much harder it is to spot places of interest in a satellite view) where NPR can clarify visualizations; I am not sure how we would apply this to data visualization design, when there is no such thing as "photo realistic" rendering.

The authors also mention research on the effectiveness of brush stroke properties, but it is unclear to me how brush strokes, which combine color, orientation, width, length--so many features--in an undefined manner can be analyzed for effectiveness. I am not very convinced by Figure 17; the shape, thickness, and length of the "brush strokes" strike me as the main influence; in fact, I may never have known that the visualization on the left were simulated brush strokes at a first glance.