Lecture on Thursday, October 6, 2011. (Slides)


  • Required

    • Chapter 8: Data Density and Small Multiples, In The Visual Display of Quantitative Information. Tufte.
    • Chapter 2: Macro/Micro Readings, In Envisioning Information. Tufte.
    • Chapter 4: Small Multiples, In Envisioning Information. Tufte.
    • Low-Level Components of Analytic Activity in Information Visualization. Robert Amar, James Eagan, and John Stasko. InfoVis 2005 (pdf)

  • Optional


wtperl wrote:

The Amar, Eagan, and Stasko paper has a good list of tasks for analyzing data, however I think they miss an important one: transformation of the data. We've seen in class and the first assignment's data set that oftentimes raw data doesn't visualize very well. But by using logarithmic calculations, bucketing, or other transformations, the designer can turn seemingly random or useless data into an interesting visualization that emphasizes a trend or makes a point. The paper admittedly focuses more on techniques of analysis over data management, but I think transformations are definitely part of the analysis. If raw data is cluttered, a transformation is critical to enable analysis.

cmellina wrote:

@wtperl, yeah that's a good observation. I would have thought it would fall under "Compute Derived Value," but I just looked again and it turns out that that category is intended for "aggregate numeric representations," so I'd say that category is a bit mislabeled. I'd change its name to reflect that the computed value is an aggregate and add in a "Transformation" operation. I think the fact that they did not find a transformation operation indicates a bias in their population (students didn't know about transformations or when to apply them) or a bias in the surveyed datasets (transformations were not necessary to answer any questions of interest).

On another note, I think Tufte's concept of data density is cool, but somewhat superficial and incomplete. It is superficial because, if you have a table of values, you can trivially increase data density by, e.g., making the font smaller. (As an ironic aside, consider Tufte's table on p. 165 in The Visual Display of Quantitative Information. My best estimates are 21 rows, each with 6 numbers, in an area 4.5" wide and 6" tall, giving a data density of ~4.6 data entries per square inch. Thus, that graphic would fall toward the bottom of itself. A few pages earlier in discussing a distribution plot, he comments that it's a "lightweight and unscientific 4 numbers per square inch." Here he needs to eat his own dog food a bit.)

It is incomplete because I think there was not enough emphasis on what the real benefit of dense data graphics is. I see the real benefit in foregrounding emergent properties of the data. The emergent properties of the data are the broad trends that span lots of data points and tell you about an underlying process - they are typically the properties that you are trying to find when you are "looking for a pattern." By making data entries small and data density large, you are more likely able to make these patterns emerge as visual Gestalts in the graphic. So to me, its much more about how best to leverage properties of the visual system than it is about the parsimony of high data density. I think a lot of what Tufte says hits close to this, but I didn't see him exploring this aspect explicitly. Another way to put is this: it's not about making the data points small, its about making their visual interaction salient. Making data points small is definitely an avenue toward accomplishing this. I just thought Tufte needed more means-to-an-end discussion.

Lastly, sparklines and datawords are awesome. Why aren't we seeing more of them around?

jojo0808 wrote:

The Amar, Eagan, and Stasko paper was definitely an enjoyable change of pace. This paper looked at how visualizations support analysis tasks, while Tufte focuses a lot on how existing graphics can be visually improved. The paper also mentioned a bit on data-centric approaches to designing visualizations (exploratory data analysis). There definitely seems to be a number of different ways to approach the process of creating a visualization.

However, creating a visualization is just that -- a process. We "look" for patterns and trends. All of what we've seen so far seems to look at fairly low-level things like how visualizations are used or how to use visual elements to make them more clear/efficient/effective. Is there a higher level description of how to create a good graphic. I guess Jeff's lecture on exploratory data analysis today touched on this a bit: you form a question, create a visualization to answer that question, come up with more questions, and revise until you arrive at a visualization that shows off what you found out from the process.

awpharr wrote:

@jojo0808, you are definitely right about the higher level description of how to make good graphics. Using an iteratively-based process to look at and visualize data is inevitably going to lead you to some of the most interesting stories that data holds. That's why I feel Wattenberg and Viégas' Many Eyes project is so powerful. With so many different brains around the world critically analyzing and playing with the visualizations posted online, new and interesting trends will emerge from the already substantive visualizations. One set of data can tell so many different, unique stories, and having a lot of people go through this higher level process that Jeff described in class today, will ultimately lead to very precise and effective visualizations. If we can supplement this process with an understanding of what analytic tasks our visualization viewers will use for a specific image, as enumerated in the Amar, Eagan and Stasko paper, then we can have an even better understanding of how to visualize the data.

tpurtell wrote:

Geography based visualizations are an interesting phenomenon. The projection of data onto the physical space we all know (or can at least read about in a book) lends itself well to letting the viewer draw conclusions. There are natural trends for things like disease spread, local behaviors determined by laws, and other similarities driven by community practices. Then there are maps of abstract ideas when area has been used to convey importance, such as full screen presentation of all of today's news headlines. As the represented data gets more abstract, it becomes much more difficult to place quantities in a sensible two dimensional arrangement.

  • Would you ever use an area graph for exploratory analysis? It doesn't seem like a great way to understand the data, though it does seem like a good medium for communication to the casual reader.
ifc wrote:

I have a few comments on the Tufte reading. The first is a reaction to how well micro/macro lends itself to geographic data (as shown in many of his examples). Perhaps part of this is that we can immediately relate to the macro part of these visualizations so the micro parts come easier. For more general data, it seems like a good micro/macro relationship would be very difficult to produce, as you are somewhat at the mercy of the data itself. The tree-rectangle graph visualization is among the best that I have seen at this micro/macro aspect of a visualization for more general data.

Secondly I think small multiples are a useful visualization type but when they are broken up by time, they may be better represented in a small interactive visualization (ie 'play', 'pause', 'rewind', etc.). Personally it just seems that detecting trends in data over time would be much easier if they be found in this type of visualization rather than a series of pictures. On the other hand, this type of visualization would not let you see more than one time slice at a time, which might be a useful function depending on the data.

cmellina wrote:

* This comment is the result of a bug. If you are reading this thread linearly, you need to go down to read @abless's comment. This same comment is down there to, in its correct chronological place. Loop-hole much?

@abless, yes, that's exactly what I mean. You understood me correctly, except my spin was that that shrinking of the whole table is a trivial increase in data density as compared to finding a better arrangement of the data. I, too, am questioning the helpfulness of the definition in that regard.

Yeah, the significance by vision idea was really cool. I suspect that over many such tests, people will be unable to detect significance although it is there (false negative) more often than they will detect significance when it isn't (false positive). So I'm betting that the human perception significance test has lower statistical power than formal tests. But that's just my intuitions about the sort of randomized arrays we saw in class today and my own experience. I'm interested in whether anyone has tested this. Anyone know? I look forward to exploring that topic a bit when some free time comes my way.

My second hypothesis is that the task will get much more difficult if you are shooting for anything more than p < 0.05. That p-value requires at least 20 plots. If you want to have p < 0.01, you need 100 plots in the lineup. Soon the task is overwhelming or too time consuming. I'm betting a cool interface could be designed to aid people in rapidly doing a large visual significance test, maybe by aiding in pairwise comparisons of plots or something. At any rate, it's not that practical, but from a philosophical point of view, it's a really cool idea to anchor significance testing to human perception.

rc8138 wrote:

At first sight, I was not able to draw connections between exploratory data analysis and the taxonomy proposed in the Amar, Eagan, and Stasko paper. My initial reaction was that many of the lower-level tasks (e.g. retrieve value, filter, compute derived value) can be performed easily (perhaps all at once) if the data are stored in database: SQL like languages would be the perfect tool for such data retrieval/collection.

With a little bit more thoughts, I realized that many data are in fact not structurally available in databases, and many of the "data pre-processing" are helpful, if not required. When performing Exploratory data analysis, one of the important tasks is to iteratively ask new questions from the visualization. By decomposing the questions we have in mind into lower-level tasks (like the ones Amar, Eagan, and Stasko proposed), we have a clearer picture in tackling the picture. More so, we will have a better idea on what type of data visualization are needed to answer and complete each of the lower-level tasks.

Such decomposition is very effective in helping us to answer the questions in mind, and will help us to utilize visualization more effectively and efficiently.

grnstrnd wrote:

As are some others, I am a bit skeptical of Tufte's glee over high-density visualizations. First, I believe that this is a solid concept--certainly graphics with super low information density are often times not worth the graphics--but almost all of his super effective high-density graphics are maps. This leads me to wonder how universal his definition is and if it shouldn't just be, "Ditch any graphic that's not a map, duh." On the other hand, even this might be a good suggestion. Certainly maps seem to lend themselves to the most powerful revelations and represent other of Tufte's principles as well, namely micro/macro. Maps are meant for exploration, so perhaps the challenge is to find a way to think of visualizations as maps; i. e. to represent the data in a way that encourages multiple levels of focus, etc. Ultimately, these seem to be the most transparent representations, plotting almost nothing except the data itself (therefore wonderfully dense), and they also lend to the most beautiful revelations on the part of the user.

jneid wrote:

Since the lecture topic today was Exploratory Data Analysis, I was trying to connect the Tufte readings with this theme. At first, it seemed like Tufte's principles would make it harder to explore the data, but I realized that when implemented well, they would allow superior exploration. Using high data density at first seems to clutter the graphic, allowing little actual data to be picked out. When exploring data, however, this allows the big picture of the data to be seen, so that overall general patterns, as well as anomalies, may be picked out. The visualization may then be drilled down to include only the data that was found interesting. In the same way macro/micro images may first seem overwhelming but allow for both broad and in-depth analysis. Small multiples also gives an overall summary of the data as well as more precise snapshots, all in one image. The challenge then, is getting past the overwhelming nature of visualizations packed with information to make them useful for exploration. To this end, I definitely agree with ifc that interactivity is the way to go. This allows the viewer to explore data on a large scale and then drill down to an interesting view that zooms in on part of the data, a microcosm, or one of a set of small multiples. Animation would also increase data density, allow zooming to macro and micro levels, and give the appearance of many small multiples using an additional dimension: time.

abless wrote:

@cmellina, I am not quite sure if I understand you correctly. Clearly, varying the font size does not change the numerator in the ratio that yields data density (the size of the data matrix), so I assume you mean to make the overall graph area smaller by decreasing the font size. But then you end up with a smaller visualization, so the "density" (of the data) increases - which is rather intuitive, don't you think?

I think the real question is whether that definition of data density is really helpful. Having a small visualization with great data density might be hard to decode for the viewer. I am not sure whether Tufte considers the role of aesthetics at all. In fact, it seems to me that Tufte is overly concerned with quantifying the quality of visualization by numbers. Data density is one example, but we have also seen "Lie Factor" and "Data-ink ratio". While I appreciate those definitions and numbers as a good heuristic, I hesitate to take them too important.

Lastly, a quick word about today's lecture: I thought it was really cool to assess statistical significance by visual perception ("find the real data"). I wonder whether this can actually be made admissible in science (given that you have enough people spotting the real data). After all, you could argue that our visual perception is a model in a similar way that that the Gaussian distribution is a model. I would be interested in hearing more about this. And Benford's Law is super cool!

luyota wrote:

I feel the paper from Robert Amar, James Eagan, and John Stasko has defined a good set of rules to apply when trying to propose experimental questions to ask and answer when trying to analyze the raw data that are new for the visualizer. However, like cmellina mentioned in the previous posts, these rules might not be comprehensive enough to cover all the aspects, and therefore I would question whether staring from these rules is a good approach to explore the new area. But anyway, I think in some sense it provides a basic guideline. Probably we'll be able to learn more about that when doing the assignment 2.

Also I can't agree with previous posts more about the lesson taught today in the class was super interesting. The idea of having human's visual system distinguish the elements of statistically significance was very new to me, and I'm curious whether there are some theories to systematically explain what happens to the whole process. That will be really cool.

blouie wrote:

@tpurtell I agree that geographical visualizations are interesting, but wouldn't the same conclusions hold for other "scientific data visualizations," where the visualization of data tends to follow spatial preconceptions of said data? It certainly would follow along with the notion that abstract data starts to become more difficult to represent in standard two-dimensional visualizations.

@cmellina I think your second hypothesis is pretty interesting, especially since findings with p > 0.05 aren't statistically significant. So basically, your hypothesis implies that for anything to be statistically significant, it requires a large number of plots. This is, in retrospect, pretty obvious, but it's nice to think about why that happens.

zgalant wrote:

I just wanted to comment that the sparklines Tufte describes are amazingly effective. It's amazing how much information you can fit in just one of those lines, and Tufte showed some great examples of how to show even more data by highlighting ranges, averages, highs, lows, and start/end points.

The baseball season graphic was one of the best. It really captured the season in a half page graphic better than anything else could. It was so easy to see the divergence between the teams and how the best ones really separated themselves from the rest of the pack.

I think these have many uses, and I especially like them used in sports and the stock market. They are surprisingly effective for their size and much easier to read than you'd expect something so small to be.

crfsanct wrote:

@cmellina. Interesting conjectures about significance by human perception. It seems like they could be related to Tufte's theory of small multiples. I think humans should generally be good at detecting significance because of why small multiples are effective, which is that humans are good at comparing similar pictures within the eye's viewing area. This is also a reason why comparing 100 plots would be difficult, since they would span a greater total area. I like how this is also related to Tufte's theory of high-resolution data graphics. The ability to detect significance is related to the ability of humans eyes to resolve and make meaning out of those small randomly placed dots in those plots. I see the overarching message that Tufte presents with his principles in that good visualizations free our eyes from doing unnecessary work and instead allow our eyes to use their natural advantages to do moremeaningful work.

cmellina wrote:

@abless, yes, that's exactly what I mean. You understood me correctly, except my spin was that that shrinking of the whole table is a trivial increase in data density as compared to finding a better arrangement of the data. I, too, am questioning the helpfulness of the definition in that regard.

Yeah, the significance by vision idea was really cool. I suspect that over many such tests, people will be unable to detect significance although it is there (false negative) more often than they will detect significance when it isn't (false positive). So I'm betting that the human perception significance test has lower statistical power than formal tests. But that's just my intuitions about the sort of randomized arrays we saw in class today and my own experience. I'm interested in whether anyone has tested this. Anyone know? I look forward to exploring that topic a bit when some free time comes my way.

My second hypothesis is that the task will get much more difficult if you are shooting for anything more than p < 0.05. That p-value requires at least 20 plots. If you want to have p < 0.01, you need 100 plots in the lineup. Soon the task is overwhelming or too time consuming. I'm betting a cool interface could be designed to aid people in rapidly doing a large visual significance test, maybe by aiding in pairwise comparisons of plots or something. At any rate, it's not that practical, but from a philosophical point of view, its a really cool idea to anchor significance testing to human perception.

bsee wrote:

I have to disagree with wtperl and cmellina that the transformation of data (like the log-transformation in the example) is not included in the paper. Personally, I think that it falls under "Characterize Distribution". Other than thinking that we are transforming the data to be on a log-scale, we can think of it as fitting the data to a log curve. We can then choose to represent the data on a log-scale or not. This flow seem to fit with the rest of the paper, as we can for example find anomalies on the log-distribution. The other interesting point is that transforming the data to see the anomalies more clearly, like the method mentioned in class, implicitly falls under "Find Anomalies". Sure, we have to transform the data from the absolute values to the residual term, but for each tasks, we are ultimately answering a question. The paper simply states that for each task, we are trying to answer a question. It does not explicitly argue the specific steps to create the visualization. Thus, it is left up to the engineer (or in this case the student) to manipulate the data to answer these questions.

That was my main take away from this paper.

The other interesting point that came up in the readings was that of Tufte's major/minor representations. When I read that, I immediately thought of how the mechanical turk example (showed in class) can immediately benefit from this major/minor representations. For instance, on a major scale, we can see the clustering of the people who didn't finish, somewhat finish, and completely finished. We can then zoom in to the different categories on why the users reached that result. Did majority of the users who did not finish experience a software bug? Was latency a reason why users gave up half way? The ability to see the reasons on a minor scale will definitely be beneficial to creating a "better" mechanical turk.

jhlau wrote:

Tufte's assertion that "Data graphics should often be based on large rather than small data matrices and have a high rather than low data density" is a statement that I think is a little (perhaps purposefully) too simple (168). I think that data exploration should certainly tend towards high density - I can't imagine a low density graphic being very useful for exploration. On the other hand, I feel that graphics meant to persuade as opposed to explore can do well in either low or high density. While high density graphics encode more information and can potentially be more persuasive for that reason, low density graphics strongly emphasize main facts. To me, low density graphics excel at summarizing - which is why they're great for persuasive purposes, but not so good for exploration.

@cmellina I also find the concept of visual significance interesting. I do think, however, that the visualization makes a big difference. It's been shown that humans recognize some things extremely well (such as faces or things that look like faces), and even today in class it seemed that some types of visualizations were more easily identified to be significantly different than others. I'd also be interested in studying this general effect through crowdsourcing, and seeing how accurate large numbers of people are at discerning significant differences.

Also, one question I have about small multiples: how large of a difference does the arrangement of the small multiples make in perceiving the visualization? I imagine that, because our vision is so relative (optical illusions, anyone?), certain arrangements might obscure data. This might go well with a Mechanical Turk experiment. You could perform the experiment by testing how certain arrangements of small multiples affect the perception of visual significance on a wide number of people.

jkeeshin wrote:

1) I agree with Zach Galant, who mentioned the surprising effectiveness of sparkling graphics in transmitting lots of binary data, but also with only small visual changes (underline, coloring), were able to transmit several more variables.

2) I felt a little bad when I read the quote from Tufte basically saying he thought "pie charts were stupid," because my first graphic was a pie chart. I thought it was a simple way to convey information and show it over time.

"A table is nearly always better than a dumb pie chart; the only worse design than a pie chart is several of them, for then the viewer is asked to compare quantities located in spatial disarray both within and between pies… pie charts should never be used." (178)

Wow, Tufte, harsh, and I guess I messed up assignment one.

But then I continued to read Tufte's "Envisioning Information," especially the chapter on small multiples. And when reading here, I stopped feeling as bad, because it seemed that small multiples had a little chance to redeem the evil that is 'many pie charts.'

"At the heart of quantitative reasoning is a single question: Compared to what? Small multiple designs, multivariate and data bountiful, answer directly by visually enforcing comparison of changes, of the differences among objects, of the scope of alternatives." (67)

So while pie charts may not have been the best way, I still contend that this design is in the feel of small multiples, and does easily allow the viewer to pick up on change.

emjaykim wrote:

Tufte's bias toward high data density seems in part to have come from his chosen media of (literal) ink and paper. Now that lots of visualizations are made on screens, have fewer limits regarding available space area (although there still is the concept of the "fold", scrolling is just so convenient), and unlimited "ink" available, the idea of space as a limited resource might weaken during our times.

One of the first examples shown in class- the facebook edge graph, with the flawed data, demonstrated that, in exploratory data analysis, human insights and background knowledge is crucial for recognizing problems and patterns.

mkanne wrote:

Like @emjaykin above I also took issue with Tufte's idea that high data density is almost always better. Going back to last class' readings, rigorous erasing of non-data ink did improve some of his examples but, in my opinion, hurt others. In Chapter 4, the 6 ways that the data was being represented in that bar chart appeared to be a way of visually reinforcing the qualities of the data. Does this not constitute higher data density? Must higher data density only include unique data points?

I would argue that for some situations like micro-macro graphics redundant data representations can be very useful. Indiscriminate (and even supposedly discriminate) erasing of "extraneous" graphical ink would cause contextual problems as the amount of data being represented rises. In the cases of micro-macro and small multiples graphics, redundant ink, even data ink, can provide a very useful baseline for the significant changes that are attempting to be highlighted by the visualization.

Although I see what @emjaykim is saying above about space being cheaper when visualizations are digital, I do see some merit in the continuing to apply the Shrink Principle to the digital realm. Humans are capable of reading very fine print on a webpage so seeing thousands of fine data points on a digital graphic is not a problem. In fact, I worry that giving digital graphic designers the freedom to think beyond the restrictions of paper size (and hence the size of what we can visually perceive) could lead to incomprehensible visualizations.

chanind wrote:

Tufte brings up some interesting points regarding the human eye's ability to distinguish intricate features in a visualization. While I think this is an important thing to be aware of when designing visualizations, I don't agree that it should be a universal theory that "higher density = better". This would be equivalent to saying: Novels are better than billboards because there are more words / area. High data density graphics definitely have their place, but so do low data density graphics. If the goal is to make a point with the least amount of thought required on the part of the viewer to understand your point, then low data-density graphics are clearly better. These are the "Billboards" of the visualization world, and they're not inherently inferior to high density graphics - they just serve a difference purpose.

pcish wrote:

The Engineering Statistics Handbook website provides a more concrete definition for the (fuzzy) concept of exploratory data analysis. Yet, the definition it offers has an emphasis on deriving a statistical/probalistic model that fits the data well. According to the site, the main difference of the exploratory data analysis approach is to analyze the data in order to find a good model instead of guessing a model and then testing the data to confirm the guess. It futher goes on to describe what constitutes a good fit (i.e. looking for characteristics such as behaving like random drawings). This definition seems to be at odds with some of the material covered in class. For example, in class it was mentioned that exploratory data analysis is an iterative process of answering questions, and unless the final question we ask is what model fits the data best, the two definitions would not converge. Also, the definition given on the site appears to lack the iterative component: one analyzes the data and comes up with the best model, there is no need to iterate. I feel that the in class interpretation is more general and matches my conception of what exploratory data analysis should be better, but it is also more abstract and harder to grasp.

junjie87 wrote:

I was pleasantly surprised to read the chapter on micro/macro readings because I think Tufte really described what I was trying to say in my previous comment about amateur/expert visualizations, where a cursory read of the figure gives a general idea, but the visualization also contains enough details for those who knows/wishes to find more information. I did not think the maps were a good example of micro/macro diagrams, however, as it seems really difficult for a person to pinpoint exactly which pixel he/she belongs to. The stemplots, on the other hand, were simply beautiful.

I was also surprised when it was mentioned in lecture that Benford's law is admissible in court, because as far as I understand nobody really knows WHY the law happens, only that it appears everywhere. This was in stark contrast to a previous news article I read where a UK court rejected Bayes Theorem (a true, provable math theorem).

stubbs wrote:

While the Amar/Eagan/Stasko shift from representational primacy (e.g., semi-quantitative analysis of the spatial substrate or a priori design principles) to ostensible analytic primacy is refreshing, their decomposition of analytical tasks is reductionist beyond the point of utility for the designer.

If we are to subscribe to Tukey's notion that a good visualization implicitly "enumerates potential root causes" and Wattenberg's notion that a good visualization "shows a problem in your data", then it follows that the taxonomic codifying of analytic tasks should be focused on a more exploratory and domain-specific mid-level tasks. Indeed, Amar/Eagan/Stasko acknowledge these fundamental flaws; 1. predicating their model on a 'student corpus' of deductive questions, rather than domain expert use-cases (or logic theory) and 2. considering only static goals in their task mapping. Only by focusing on mid-level tasks, can we begin to weight the significance and consider the trade offs of each of said task's component lower-level facilities in spatial and temporal (i.e., viewer cognitive progression) significance and illuminate the domain-specific utility of each visual element.

As opposed to a rote low-level checklist, mid-level tasks naturally give rise to interesting measures of exploratory efficacy (e.g., mutability/clarity ratio, procedural/spatial congruence, etc.) which seem more relevant in the modern interactive/social Many Eyes paradigm. It's the elevation from, "did I remember to put a steering wheel on this car?" to "how will this car handle at Laguna Seca in the rain?"

Question: implementation details of stats methods are critical in a research setting, however, do you think that many stats methods are gleefully/purposefully framed as abstruse (and often redundant in essential function)? Is there a demand for an easily-visualized high-level (e.g., not R/Matlab) stats web framework for the general public? Does this exist?

kchen12 wrote:

Tufte's statement that clutter and confusion are the result of bad design and not the nature of the information itself was a great clarification, further emphasizing our responsibility as visual/information designers to our readers to clean/prime our data and create many rich levels of interpretation per visualization.

What Tufte says in Micro/Macro Readings about high-density information initially contrasted with what I had gleaned from the earlier chapters of The Visual Display of Quantitative Information, where he showed examples of reducing bar charts to mere lines and quartile plots to dots. As someone who does graphic design for a hobby, I've always focused on using white space to my advantage, trying to be clean if not minimalist, and often trying to strip away as many numbers as possible in my infographics to make them low density, which is what I thought Tufte was encouraging in those chapters. After some thought however, I realized that high-density information and high-density ink are two different concepts--thin data does not equal a thin amount of data ink--and that Tufte has different expectations for the richness of data and the richness of a visual image.

In graphic design nowadays, I've seen tall bold fonts like League Gothic are often utilized, and it is a common trend to see titles in all caps. However, Josef Albers is quoted on page 51 saying that because of equal height, volume, and width, only capital letters present the most difficult reading. It seems that visual designers nowadays overfocus on aesthetic versus interpretation, and I wonder if this is a trend that will continue, moving away from what Tufte advocates for information visualization.

I was introduced to the concept of sparklines this summer when talking to medical surgeons for a project I was working on. They asked for some way of visualizing changes in heart rate from regular to irregular rhythms as well as conveying a drop or increase in beats per minute. It is cool to see that they are an actual design concept as described by Tufte of encoding a significant amount of data in one word-graphic, in comparison with a set of numbers or clunky scatter plot.

kpoppen wrote:

@stubbs I'm not sure I understand how Amar et al. is reductionist beyond a point that is useful to the designer. While not a designer myself, it seems to me that having a taxonomy of the analytical tasks that visualizations are often created to address could be very useful as a reminder to ensure that the visualization is well-suited to the kinds of comparison that the designer was hoping to address. Even though the author admits that the list is incomplete because it was created from student questions, I don't think that that really makes the items that the authors have identified any less valid.

That said, I would agree that at the low-level that the authors address a lot more of the value would be in combining their taxonomy with an analysis of what kinds of data visualization best address each task (as others above have also suggested). The combination of these two things would be very useful, especially for things like automatically generating visualizations that address certain kinds of analytical tasks (e.g. "present this data so that I can see the ranges well, and to make clustering on parameters x and y as clear as possible", or something).

zhenghao wrote:

It was very cool to learn about the analytic task taxonomy that Amar and colleagues designed especially since I wrote in my previous discussion post that it seemed like a desirable alternative to the data based taxonomy :)

That said, I wonder if there exists a database or list of visualizations which best support these tasks. It would be a great resource to just put in a set of analytic operations and get a suggested list of visualizations. Sort of the kind of ranking Bertin or Mackinlay came up with for the data taxonomy.

Also, just a comment on the log transform discussion, it seems like Amar et al. were interested in classifying the kind of analytic operations the reader would be interested in performing and every one of the steps one would perform when analyzing the data. E.g. operations such as imputing missing values are also similarly omitted although they are important parts of data analysis.

The way I see it, the log transform (or any data transformation for that matter) is a computation that we might perform to support certain analytic tasks (as defined by Amar et al.)

For example, if the question were, does quantity A grow exponentially with quantity B? We might perform a log transform on A such that an exponential trend becomes a linear one (which is more perceptually detectable). But the task in this case is to correlate. Or if we do a log transform to compare 2 quantities which vary across many orders of magnitude, we're still supporting the correlate analytic task. Even if the task was to figure out the what the log of quantity A is, the log transform is just a computation used to facilitate the retrieve data task (which allows the reader to just read the value off a provided axis instead of reading the raw value and taking the log).

The authors seemed to have anticipated some of our objections and it seems that data transformations might fall under the category of low level mathematical actions mentioned in the section on "Omissions from the taxonomy".

joshuav wrote:

@turptell I think that using a map for exploratory analysis seems cumbersome and difficult with regard to data cleaning and actual visualization creation. If geographic meaning is a concern, multiple bar graphs (assuming uni-variate data) showing different regions of the world would be easier to produce and likely show the same trends.

jessyue wrote:

I disagree with emjaykim's point that Tufte's bias toward high data density comes from the choice of media. Although on printed forms, the amount of space is a big consideration, this is not to say that in digital formats, space is unlimited. The amount of space that human can view at one time is always limited, and screen real estate is always valuable. I believe Tufte does provide a valid point on packing high density data in a visualization regardless of the medium it is presented in. At the same time, I believe this principle needs not to be blindly applied. As always, one needs to look into the audience and the message to be conveyed. If the audience is the general public, and the message is simple and needs to be conveyed in a short glimpse (such as on a highway billboard), I believe for both "data-ink" and "non-data-ink", less is more. However, say the audience is a scientific group, and the message being conveyed is some complex finding presented on a scientific poster/paper, Tufte's advice of packing more data in less space does give the overall visualization more richness, authenticity, and performance.

jsadler wrote:

To Small Multiple on Not to Small Multiple ? ... Unleashing the MotherF**king Moonwalk

What is a better way to understand the Michael Jackson moon walk ? Through a small multiple like this article on unleashing the motherf**king moonwalk OR through a youtube video like this ?

Tufte show some great examples of using small multiples effectively , for example the LA air pollution visualization (pg 168 TVDQI) clearly shows a growing geographic trend. When it comes to static printed paper I love this technique BUT with dynamic digital displays we are no longer limited to static graphs of the past.

@Ifc and @emjaykim hint nicely at the short comings of small multiples contrasted in a digital world - with unlimited ink jumping alive with revealing data movements. Tufte himself shows some cool breakdowns of complex "dance steps" (which i personally find very difficult to decode) and hints at the need for a clearer "dance notation".

I would argue that an interactive graph animated over a time sequence is generally a way better way to show more complex time sensitive data changing over time. Just recall the revealing ripples shown in the animated population pyramid show in an earlier class.

But when is a better time to use a small multiple vs an animated graphic ? Sometimes you of course only have physical paper to work with. Sometimes you just don't have the time to watch a video. .Sometimes you really want "frozen frames" to see an exact point in time - a great example of this is the small multiple squence of The Horse in Motion that proved that horses indeed have all four hoofs off the ground during a gallop.

Does anyone know of existing research to quantify the visualization quality differences between using a small multiple vs animated graph ??

In the mean time i'll be practicing my moonwalk with youtube....

aliptsey wrote:

@cmellina I agree that the comparisons between human perception and statistical significance testing that we looked at in class were pretty incredible, but that false negatives were also a possibility. I think that a really interesting way to approach this problem might be not only looking at the recognition (true or false) but also looking at the time required to make that assertion. It might be the case that attributes of images or data visualizations are perceived more quickly or to a higher degree of accuracy than statistical significance tests, but the difference might only be revealed when looking at response times. If anyone has any papers/studies to recommend on this topic I would be really interested to get some references to explore this more.

The perceptual ability of humans is yet another reason tools like Many Eyes are so powerful and exciting - you can pick up patterns or develop interesting observations very quickly, and then use statistical methods to test your data rigorously.

yangyh wrote:

The analytical tast taxonomy proposed by Amar, Eagan, and Stasko is definitely useful as a guideline for us when doing data analysis. However, to me the taxonomy seems a little bit too fuzzy - it would be better if the tasks categorized came up with a set of suggestions on which visualization should be used to most effectively and expressively the task. That said, following the guideline surely helps in my assignment 2, and I hope I can experience more about it when doing the assignment.

Also, the demo showed in class yesterday was awesome, especially the "read data finding" part. I was also impressed by the quote by Martin Wattenberg. Indeed, sometimes what we only care is the correctness of the data because we're too confident and dependent on laws of science, or even our own judgement. We should always keep in mind that problems discovered might actually be the most precious information in the data, and should never feel sad about them - because they might help us find what the truth is.

vulcan wrote:

I found the Tufte chaper on High Resolution Data Graphics very helpful, and will definitely keep some of the techniques in mind.

In particular I liked the technique of "small multiples" to visualize time series data with multiple variables. Instead of a typical graph with time on the axis and the variables on the y axis, small multiples allows one to succinctly convey the change of multiple variables over time, especially in the spatial domain.

However, one potential issue I see over this is that by presenting information in discrete time steps, it lends itself to misrepresentation/omission of data. For example, depending on the "sampling interval", critical moments in time may be overemphasized or skipped over. So to use this technique, one might have to cycle through many time intervals/offsets in order to present something meaningful to the viewer, and also be mindful of potential misrepresentation.

yeyleo wrote:

I'd like to talk about a point that was brought up in lecture, well summarized by the Wikipedia article on exploratory data analysis: "Tukey held that too much emphasis in statistics was placed on statistical hypothesis testing (confirmatory data analysis); more emphasis needed to be placed on using data to suggest hypotheses to test."

While I obviously believe that data visualization is a powerful tool in helping us to understand datasets, I'm also a bit worried that one can manipulate visualizations to show trends that might not be in the data (similar to how statistics alone cannot detect the difference between the 4 datasets we saw in lecture). By stretching an axis, or using a clever visualization, I think data can be made to show something that it does not really represent. I don't think that this necessarily has to be malicious; rather I might use a data visualization that leads me to conclude something about the data that it does not really support. My questions are: - how do we know that the data visualization we have chosen are correct and actually do tell the story that our data is trying to show and not something else? - how can we be sure that a visualization we see is actually supporting the conclusion reached by an author?

It seems that only a combination of statistical tests and visualization can lead us to say that certain data supports certain hypotheses. But if one is not sure which direction to go with a large dataset, how can he or she pick the correct visualization to use?

bbunge wrote:

"Clutter and confusion are failures of design, not attributes of information." -Tufte

In Micro/Macro, Tufte presents images that are created by combining many small pieces into a whole. Rather than removing detail for clarity, this technique adds detail to create images that show the whole as the sum of its parts. The result provides overview as well as detail in the same image.

In Small Multiples, we again see the pattern of providing an overview with detail. Each image presents a dimension of the data. This dimension can be an aspect of the model or a point in time among a sequence. Through repetition, the variations become visible at a glance.

Both of these methods get me thinking about how animation can be used as a tool to aid the message of the visualizations. I have seen visualizations that offer zooming in for more detail and ones that display small multiples as frames in an animation. I venture to say that neither of these "improvements" add value to the visualization. In light of the Tufte readings, I'd venture to say that they remove value. Zooming in prevents viewing each piece in relation to the whole. Viewing small multiples one by one disregards the purpose and advantage of viewing all at once to make comparisons.

Are there ways in which animation adds substantial value to the visualization?

schneibe wrote:

The macro / micro chapter made a big impression on me. I think that this kind of techniques should be used more often. As demonstrated in class, half transparent scatterplots could (and maybe should) be overlayed other representation such as boxplots or histograms. This would avoid the problem of not seeing outliers or anormal distribution. More complex modeling techniques (e.g. path analysis, mediation analysis, multilevel analysis) could include more details on the data by adding small multiples to the standard graphs; this could be done without increasing the cognitive load of the user. The challenge is to provide representations of complex statistical procedures that even novices can understand with minimal explanations.

bgeorges wrote:

I enjoyed the (brief) discussion of Benford's law at the end of class. For anyone following what has been going on with Europe's economy, this article from earlier this year is very interesting (you need to be on campus to access it freely). The authors took macroeconomic data published by submitted by the European governments to the central statistics database and then measured the deviation of the distributions of first digits in this data from the distribution predicted by Benford's law. Greece (which is now known to have manipulated economic data) came in first, so it definitely shows that this model is effective. The interesting thing is that Belgium, which has thus far not been discussed as a major risk, tended to have economic data that deviated from Benford's distribution. In light of this, perhaps it should not be so surprising that news came out today that Belgium's credit rating might be lowered. Here's a screengrab from the article for anyone that's interested: {{j1468-0475201100542xpdf_%28page_9_of_13%29-20111007-164151}}

dsmith2 wrote:

In response to bbunge's open question:

Tufte's examples of micro/macro do not in any way suggest that animation is not an important feature and cannot "improve" visualizations. Some visualizations do not need animation, but others greatly benefit from animation.

You can show a person one picture, and then another, and the differences may be clear, but it would break one of the primary principles of visualization to hope that people can do the interpolation in their head when animation can physically show how the data changes from frame A to B! It would be a crime not to offload that data in visualizations of discrete time intervals for instance. There are many more examples, especially when you are visualizing things that have a changing physical form.

Additionally, although Tufte's examples show how viewing small instance data in a larger context can provide huge insight, it really depends on your data. The interesting patterns in data may be on multiple levels, so looking at information at too extreme of any scale could potentially cover up jewels within the data.

jofo wrote:

Prof. Heer mentioned something about how the field of exploratory data analysis was drawn from (or required) knowledge in databases, statistics and HCI (please correct me if I'm wrong). I like to see the contributions of these fields form a toolbox more useful than the sum of its parts. Amar et al however, did only briefly touch on how their taxonomy was more useful in combination with visualization than as basis for SQL quiries. I would have liked to see more advanced questions as well, from "domain experts" analyzing more complicated data.

On another note. From todays Tufte reading I learned three useful statements:

  • - list numbers in a way that also visually tells something about the data (Japanese timetable example) - alphabetical ordered data means missed opportunity to show more of the data - "Comparisions must be enforced within the scope of the eyespan"
netj wrote:

I had a similar concern with misleading visualizations as @yeyleo suggested, but I believe it becomes clearer if we think the main role of visualization is rather to leverage our cognition to enable effective exploration of new ideas from data, than to rigorously verify a hypothesis using our eyes. Although the significance test by vision shown in class was very interesting, I think our perception is generally vulnerable to distortions as we've seen in many of Tufte's chartjunk examples, so it does not have the potential to replace any kind of statistical tests. It is definitely possible to present false impressions of data either with malicious changes or by mistakes, but I believe this can also happen in statistical analyses too. A good way to workaround this problem for visualization seems to be incorporating more data into it, as Tufte suggests to increase the data density. I see statistical tests and visualization as complementing each other, so one needs not pick the correct visualization at the beginning. Instead, he/she can try drawing various high-density data graphics to find a good hypothesis, which can be later verified using statistical or mathematical methods. This view, in my opinion, aligns well with Tukey's emphasis on methods that suggest hypotheses.

ardakara wrote:

I also want to take a moment to talk about sparklines and datawords. As mentioned above, I also think these are tremendously powerful tools, and should be used more often.

@kchen12 I'm happy to hear at least they are used in the medical circles. I'm yet to see them used (in their line fitting form) in any publication, and I agree with Tufte that they are pretty obvious in what they mean, their learning curve is pretty flat for a new word with so much expressive power.

I was initially surprised by Tufte's excitement about using high resolution graphics to push the data density limits of our vision. However, I soon realized the distinction with "For non-data-ink, less is more. For data-ink, less is a bore." I don't agree with the first part as much, since the dot-dash-plot is very hard to make sense of at first sight, but I definitely agree with the second part, as long as the data is still kept accessible at a higher level, with extra attention to detail and investigation producing more information about it. So, I think the high density chart should be made such that it should still be useful at a glance, but unfold lower level detail upon closer inspection.

chamals wrote:

After reading reading Tufte's chapter on Micro/Macro Readings, I realized how gorgeous data can really be. Seriously, some of those visualizations were awesome. I have to agree with @cmellina comment that "its much more about how best to leverage properties of the visual system than it is about the parsimony of high data density." A high-density is not enough. Although in such visualizations you want the reader to draw their own conclusions about the data, there must be something there that can be discovered.

On a different note, I really enjoyed the section on the Vietnam Memorial. I remember bits and pieces of the story behind it and I have seen it multiple times, but I never really considered the design choices described in the reading. The design choices for art data visualizations take into very different considerations than visualizations where you try to gain insights from the data set. I think it is incredible the stories people are able to tell when they create such artworks from such data.

elopez1 wrote:

I agree with a couple of main criticisms others have had towards the Amar, Eagan, and Stasko paper. Two main points that others have made are

1. Amar et al. forgot one main task for analyzing data: that of transforming the data to better see trends

2. Data analysis should be iterative in nature. Thus, simply coming up with questions and analyzing them is not enough.

Perhaps a combination of Exploratory Data Analysis and the tasks described by Amar et al. would be effective: treating data analysis as a process where we come up with more and more questions as we go along and analyze them in the framework proposed by Amar et al.

fcai10 wrote:

I found the Micro/Macro chapter very interesting, especially a point that is made early on: "This fine texture of exquisite detail leads to personal micro-readings, individual stories about that data...". The idea that a data visualization can be simultaneously personal and general, that the individual data points together provide their own context -- such as the mesh maps, and the Vietnam memorial... this idea was a revelation to me. One of my favorite graphics is the elegant Japanese timetables that employ the stem-leaf design -- the actual numbers that represent the minute of arrival form bar graphs of the frequency of trains by the hour.

This reading got me thinking about what types of data could benefit from a visualization that provides micro/macro readings. One possibility is medical data. There are a lot of data points in the form of measurements, yes/no answers, etc. but there is also a natural context for them -- the body. While medical histories and blood work charts may be very usable, I wonder if there is a better way to visualize the a patient's data -- perhaps to save the doctor page-flipping time, to facilitate diagnoses. Obviously, this is a hefty question to address, and I'm afraid I do not have a satisfactory answer at the moment. =P

angelx wrote:

It is interesting how very powerful the human perception system is, and it makes sense to draw upon the power of the ability of humans to detect visual patterns. In general, I would agree with Tufte's point that it is desirable to present a high density of data and thus allowing the human visual system to make sense of data that would be otherwise difficult if presented as a table of numbers.

At the same time, I'm not sure having low data density is necessarily bad. For one, data density alone does not make for a good graphics as it would be simple to decrease the size of an image indefinitely and increasing the data density arbitrarily, beyond the when the human eye can effectively pick apart the details. Sometimes there simply may also not be that much data to display, in which case, I do not see a need to artificially increase the data density. In addition, simplification of the data can aid the eye and enable the reader to have a more useful mental representation (road maps are a good example of when the simplified 2d representation may be more useful for some purposes than depicting where every rock or tree is).

mbarrien wrote:

One of the things I notice missing in the discussion here about animation vs small multiples is something that Tufte stated after the maps of China: "Comparisons must be enforced within the scope of eyespan." In an animation, you're forced to look at the comparisons in linear ordering; doing comparisons between things that aren't subsequent frames are in effect not within eyespan.

So for example, in the LA pollution map, comparing specific pollution sources from the morning vs. the afternoon is lost in animation. You can kind of see a hotspot evolve if you know what to look for in the animation, but with a static page with a small multiple graphic, the comparison is quick flick of the eye, rather than a manipulation of a video player to jump between specific frames. Or more likely, jumping back *and forth several times* between frames.

Also our eyes can be quite blind to changes within an animation that happen gradually even to seemingly obvious changes of hue. These animation related change blindnesses are hard to hide in a static page. See this for an example: http://www.newscientist.com/blogs/nstv/2011/06/friday-illusion-can-you-spot-the-change.html

babchick wrote:

I agree with the opinion of @emjaykim that Tufte's catch-all suggestion that high information density graphics should be preferred is not entirely relevant today. As previously mentioned, we no longer have the same restrictions on ink, and even more interesting, I think there is much greater variance in the purpose and audience whom we make visualisations for today. I think Tufte's visualisation utility function focuses on artists, dataheads, and the immediate stakeholders in the data, but ignores the prospect of using visualisations for persuasive or educational purposes, in which case I can easily imagine low information density graphics being more effective.

I also think that small multiples is an under-appreciated visualization technique for complex data sets that span many variables over time. Especially for maps and spatial data, small multiples allow us to make comparisons of the macro structure of data over time. Line charts are great when we need to make comparisons within a relatively narrow time interval around our data points, but small multiples enable arbitrary before-and-after comparisons without losing much of the benefit of a line chart, if both are viable options.

phillish wrote:

@bsee: I agree that enabling interactivity on a visualization immediately opens up many opportunities to improve on its clarity and function. For instance, we can omit excess labels and focus the user on important stats from a glance. If the reader wants more information, they can then interact with some visual element to zoom or expand to reveal more information. The mechanical turk examples shown in the slides are static 2d visualizations, which require a balance of presenting an appropriate amount of data while reducing clutter to keep the data readable.

What interests me most in the paper is the formal definition of "high-" and "low-level" data-driven questions and their taxonomical breakdown into individual tasks. While the concept of reorganizing a complex question ("Sort the cereal manufacturers by average fat content") into separate, common tasks ("Compute average fat" and "Sort results") is an intuitive one, the paper introduces the proper vocabulary to better understand and document it.

insunj wrote:

In Amar, Eagan, Stasko's paper, I found their statement on " We argue that a stronger focus on user tasks and analytic activities in information visualization is necessary as current tools do not seem to support analytic activity consistent" to be very interesting. Although there are some standards and common methods in data visualizing tools, the paper was saying that in order to better aid analysis as human activity we should focus more on what is needed and important in that analysis pace. I found this very intriguing because it seemed this is how Minrad came up with sensational visualization then!

daniel89 wrote:

I'd actually like to propose something slightly different- based on interactivity and "data hiding". Tufte seems to suggest that high density graphics are desirable rather than low density graphics. Yet this is within a tension system where low density graphics are "instantly understandable", while high density graphics require a closer study and understanding of the representation.

Perhaps a interactive visualization that "adds data" at each frame may be a good idea. Starting off with the simplest illustration, it animates it to "add data" and explain the new dimension. Humans have a limited cognitive ability, but we can overcome this incrementally. Building up the data density incrementally may allow us to overcome this limit.

stojanik wrote:

I thoroughly enjoyed the reading on micro/macro data exploration, especially the section about the Vietnam Veterans Memorial. I agree with both @fcai10 and @chamals that it represents a much more sophisticated piece of architecture than it first appears (to me anyway). I visited the site a couple of years ago and at that time it did feel very understated and I remember thinking that their sacrifices were deserving of a more ornate memorial. But it was, and continues to be the engagement of the tourists/families/participants, moving through that shared space together as "colleagues" rather than "interruptions at an architectural performance" (p.44) that resonates. It catalyzed a lot of ideas for me about movement/pacing, transition, grouping, surface composition, and positioning, shared experience and how those concepts can be applied to immersive visualization-exploration. The micro/macro reading of the physical space is analogous to the data space whereby "Panorama, vista, and prospect deliver to viewers the freedom of choice that derives from an overview, a capacity to compare and sort through detail. And that micro-information, like smaller texture in landscape perception, provides credible refuge where the pace of visualization is condensed, slowed, and personalized". (p.38)

After looking at the Vietnam Veterans Memorial image I was reminded of a similar images, a similar perspective portrayed as a live data wall and immersive film exhibit in New York. Although they signify different things the space-process of exploration(micro/macro) was remarkably similar.

This reminded me of Kubrick's Monolith from 2001: A Space Odyssey:

pmpai wrote:

The visualization dealing with the different bacterial reactions to antibiotics was extremely interesting. That example helped to explain the value of visualization in pointing out possible faults and mistakes in assumptions. To see how the one family outlier was eventually proven to be from another family was astounding. Such applications could make so much more sense in today's public education system, patient care, policy discussions and in many more areas. This would make data hiding far more difficult as well. I'm looking forward to using these principles in the future.

Leave a comment