current page

history

user

Lecture on Thursday, October 20, 2011. (Slides)

Readings

  • Required

    • Postmortem of an example, Bertin. (pdf)

    • Visual information seeking: Tight coupling of dynamic query filters with starfield displays, Ahlberg & Shneiderman. (html)

    • Visual exploration of time-series data, Hochheiser & Shneiderman. (pdf)

  • Optional

    • Generalized Selection via Interactive Query Relaxation. Heer, Agrawala, Willett. CHI 2008. (pdf)

    • The cognitive coprocessor architecture for interactive user interfaces. George Robertson, Stuart K. Card, and Jock D. Mackinlay. UIST 1989. (pdf)

    • Exploration of the Brain's White Matter Pathways with Dynamic Queries. Akers, Sherbondy, Mackenzie, Dougherty, Wandell. Visualization 2004. (html)

  • Demos

  • Videos

Comments

jneid wrote:

The Bertin reading was interesting, especially due to the definitions and process it describes. I wanted to comment, however, on Bertin's assumptions of what aspects of the "decision-making" process can be automated. First, Bertin claims that defining the problem, or creating a list of basic questions, is a "problem of imagination" and cannot be automated. I think it is possible, however, to define questions based on the data by looking for patterns and outliers and then asking why they appear. It seems like this basic question is the essence of every human-generated question, whether asking about correlation or change (patterns) or specific data values (outliers). Beyond this, automated processes could also take into account the type and identity of the data and use outside information from sources like the internet to refine questions to what would be interesting to the intended audience. Bertin then allows that simplification could be automated, but states that extrinsic information could not be taken into account automatically "by definition". I think there's no reason an automated process could not get the relevant information and then use it to further refine the graphic. Perhaps at the time the text was written in 1981, computers weren't advanced enough for people to contemplate the possibility of the automatic processing of data or such vast resources as today's internet. There's no physical property that prevents machines from achieving these tasks, however, and with today's computing power and advances in fields like artificial intelligence, I think it's entirely plausible to create a completely automated decision-making tool.

ardakara wrote:

I think that interaction is a great point where we can reconsider Tufte's data-ink argument, in which he proposes that the data should take the spotlight in a visualization, not the design or the presentation. It seems to me that one can come up with various complicated ways of interacting with the data, but as long as the underlying data doesn't have the detail or the complexity to give birth to exciting findings, it will be a sad day.

The examples we saw in lecture today were great examples of data overpowering design. Zipdecode offers a very simple interaction that's inviting and streamlined enough to engage the user and address primary objectives they might have. However, what ends up being the climax of the interaction is user's findings in data, not the features the tool has. The features should make the findings more easily found while staying as invisible as possible.

Zipdecode had a more recreational use and obviously one can think of contexts where more powerful data analysis techniques would be useful. However, they will still be only as exciting as the discoveries they accommodate in the data.

phillish wrote:

I wanted to talk more about Fitt's law mentioned in class. It's basically defined (on Wikipedia) as:

T = a + b*log(1 + D/W)

where: T = average time to complete a movement a = start/stop time of the device b = inherent speed of the device D = distance from the starting point to the center of the target W = width of the target along the axis of motion

What this formula essentially boils down to is that, assuming the device constants are unchangeable, designers can choose distances and sizes of buttons or other interactive elements to optimize the time T it takes to click it. In other words, a target further from the center of the screen or users cursor should be made larger to afford the user some error tolerance.

It is important to note that screen edges and corners are a special exception. They essentially have "infinite" size since the mouse will be caught by the screen borders. This is why we usually see important visual elements to reach (Start/Apple button, close/minimize/maximize buttons, etc.) at the screen corners. These buttons can afford to be visually smaller, as well, given their larger click-able area.

Window scrollbars are another great example. They can be made trivially thin, but as long as they are long along one axis and rest along a screen edge, they can easily be accessed.

chanind wrote:

The Ahlberg and Shneiderman paper presents a filter and scatterplot type interface for searching for exploring data. The benefits of this interface vs looking through a list or table of text results is obvious, but the authors fail to qualify the wisdom of their design decision with quantitative results from their user studies. Essentially, they just conclude that users "like" the program they built. As was pointed out above, several of the design decisions such as the "alphaslider" are not effective filters for searching through the data. It would be more insightful if the authors presented comparisons of the time it takes users to accomplish various searching tasks using their interface vs other existing interfaces.

stubbs wrote:

Ahlberg reminds us that as a user moves within and manipulates abstract objects through progressive refinement in a space (i.e., interaction), there must be as much consistency and constancy as possible; display invariants must be preserved (and alterations must be deliberate and immediately understandable) such that the "eye simplifies".

The TimeSearcher effectively follows this directive. I'm interested as to how more complex TimeSearcher temporal query widgets would be implemented on the interface level. Working with medical data, time queries can become very complex (as Jeff alluded to today), e.g., "return 100-day intervals where an event X occurred 3 days or less before the end of interval A and between 30-38 days before event Z". Using visualizations to facilitate understanding of massive biological data has seen varied success (for the current state of DNA viz, see: UCSC-GB), but I love the idea of using visual temporal/ordinal pattern recognition tools (i.e., TF binding sites), which also relates well to Ahlberg's mention of "fuzzy searching".

ajoyce wrote:

In the Ahlberg & Shneiderman reading, and in class, we saw the example of FilmFinder, which uses multiple tightly coupled control mechanisms to interact with a starfield display of films. The interactive paradigm of FilmFinder stands at odds to that of zipdecode and others in that it goes to great lengths to avoid requiring original textual input by the user. The pinnacle of this aversion is seen in the "Alphaslider", an alphabet-based selection mechanism not unlike the alphabetic scroll bar found on an iPhone's contact list. Unfortunately, so much emphasis is placed upon simplifying the user input requirements down to such a slider mechanism that a great degree of usability is actually lost in the process.

From this example, we can draw a larger lesson in interaction mechanisms: simplicity is good, but only if it helps clarify and focus the user's intentions. In the case of FilmFinder, the user is unlikely to intend to search for movies or actors primarily by the first letter in their names, meaning that the Alphaslider serves largely to distance the user from their intended result by obfuscating the film title and actor filters through an unhelpful selection mechanism.

kchen12 wrote:

The additional principles of direct manipulation outlined in the Ahlberg and Shneiderman paper are very close to HCI principles of good design. In Don Norman's The Design of Everyday Things, two of the design principles he introduces--visibility and feedback--are echoed in the concept of dynamism in querying. Visibility means that one can see the state of a device and all possible actions, as well as grasp how to execute functions and manipulate objects intuitively. Meanwhile feedback is ideally immediate and synchronized with user action. With dynamic querying you are constantly receiving such feedback leading to a more seamless and delightful data exploration process. The idea that actions can be rapid but also reversible gives users ample freedom, another HCI heuristic.

The time-query and query by sketch examples all seem to successfully minimize the gulf of execution, and remind me a lot of the types of novel interactions introduced with tablet and touch devices. Just as Apple's touch devices revolutionized and standardized how we think about zooming, framing, clicking, drawing digitally, etc., dragging time-query boxes and sketching trend lines to query are both new and creative ways of expressing search queries.

Finally, we've frequently read that memory is often a constraint that limits data interpretation, but I appreciated that Bertin took that statement one step further by explicitly saying that in a linear data table, relationships and similarities can only be gleaned from memory, further driving home the point of the power of interactive and data visualization.

abless wrote:

Great lecture today. My first feeling is that interactive visualization somewhat bridges the "gap" between data visualization as used to convey a message on the one hand, and data exploration on the other hand. Let me explain what I mean by that: if I want to convey a message to the viewer, that's more easily perceived visually, then I think a static image is often still the most effective way. I admit that this is not always the case, and often depends on the actual data and visualization itself, but if I want to convey a message I want to ensure that the visualization is easily decodable. With an interactive visualization, the viewer's interpretation of the data greatly depends on his interactions, and so different viewers might actually draw different interpretations from the visualization. However, it seems to me that interactive visualizations (such as the zip code and baby name visualization) serve as great tools for data exploration, while providing an easier interface than Tableau, for example. In essence, such visualizations enable novices to explore data very easily without requiring them to learn extra syntax.

grnstrnd wrote:

I think possibly the most elusive concept from our class, which Bertin emphasizes multiple times, is simplicity in visualization. This is a difficult concept because we are dealing with a simplicity that is derived or engineered from extreme complexity, and often the resulting visualizations exhibit all of the complexity of the underlying data--only transformed. Bertin's exploration was a basic example of that: a small table, transformed in space and into histograms, became simple bands of black and white over the underlying numbers. He reiterated that our eyes will simplify the images if we only reorganize the data in such a way that they can, without sacrificing information. This process, although simply phrased, is I think the most difficult that we face. (Certainly, Tufte is all about it through small multiples or macro/micro readings.) The examples we have seen vary widely in the amount of detail they actually portray: even today, the baby names demo shows much simpler data relationships than the baseball players demo. Still, both are simple and reveal information to any user in every moment of interaction.

crfsanct wrote:

It would have been nice to have read the Bertin reading before assignment 2 because it describes most of what we had to do just on a smaller scale. Even though we were dealing with larger data sets, we did in fact only have a single table (csv) to work with. Following a specific example in the reading through the five stages gives a better understanding of how to go through each step. Still, it does seem as though most people would go through that kind of process naturally.

It was also nice to read about the key concepts of dynamic query filters, startfield display, and tight coupling. These are things that I definitely take for granted since they exist in many applications today. In fact, http://www.padmapper.com operates like HomeFinder.

zgalant wrote:

It was interesting seeing the NameVoyager app in class, since it's basically a copy of a CS 106A assignment called NameSurfer. I never really liked that assignment that much, since it didn't really create very good visualizations, but it was cool to see how just a few design decisions made the exact same idea interesting. One of the biggest differences is the interactivity of the display. In NameSurfer, you just create a static display, but NameVoyager makes the data much clearer by highlighting the names and providing data on demand when hovering. It reduces clutter and makes the visualization much more appealing.

luyota wrote:

I found the paper "Tight coupling of dynamic query filters with starfield displays" from Ahlberg et al. is an interesting read, especially the tight coupling concept it mentions. We're too used to the way the software work, but sometimes we don't notice that there are really a lot of design principles involved in all components deisgn. One example from the paper is that when we're saving a document, we click on the save button and the save button will become gray and nonclickable until the save is complete. Although this seems trivial to me, after recalling the software I have used, there are still many programs that don't have this single and simple feature and will thus result in errors. Well, the save button is one that doesn't really matter when it's clicked twice, but there are more buttons controlling irrecoverable actions, and that's why this design principle is important.

I also like the output-is-input idea. I think it is kinda the fundamental element for interactive design, and it's everywhere from text input to D3. Now the world is proposing natural user interface, and I guess there will be more output-is-input ideas even when speech recognition or gesture is involved.

jojo0808 wrote:

I found it interesting that Bertin referred to information as the relationships between pieces of data, rather than the data itself. In other words, data (intrinsic info) does not equal information (extrinsic info) -- it only becomes information when it we (human beings) can make sense of it, and the format of the table didn't allow for that. Only when the data was turned into a useful visualization did it become information.

I also really liked reading about employing tight coupling to direct data exploration -- it makes so much sense and I'll keep an eye out for it as I use various interfaces. The Zipdecoder example in class was also a really fun and good example of dynamic querying -- and yes, I did often stay up at night wondering about how the zip code system worked... and now I know! Yay!

ifc wrote:

First, I enjoyed today's lecture. The efficiency vs expressiveness breakdown definitely followed my intuition about a few tools, but it was nice to have a larger picture drawn for us. Likewise I asked if anyone had a quick break down of Protovis and D3 in the comments of the last lecture, so I was pleasantly surprised by that too. And of course some of the interactive visualizations were superb.

Anyways I'm curious about the Dynamic Query tools from a performance standpoint. Some of the tools we saw are web-based and I'm not sure if all data is stored locally and 'filtered' locally too. It would be nice to have actual queries to a database but I'm not sure the visualizations would be smooth because of network latency (ie draw a box and drag -> many requeries/redraws before stop of drag).

Also some the early tools in the reading have pretty basic maps between requests and raw SQL. One of the really cool things about Tableau is that you could define your own custom metrics through an internal language. Though (I don't think) the language is translated to a SQL query, I still think neat that it gives you so much power.

Finally the Bertin reading had a number of questionable quotes. In his words, "… graphics is not an art. Its is a strict and simple system of signs".

yeyleo wrote:

I enjoyed the zip code example as well as the baby names example that we saw in class today. However, one thing that stood out to me was that both data sets lend themselves very easily to an effective visualization. It could be because both have ways that they can be sorted as well as searchable, therefore making themselves conducive to fancy visualizations such as those that D3 provides. But how would one go about making a great interactive dataset for data such as SCL levels over a day? It doesn't seem like that type of data can lend itself to a "wow" effect the same way that the zip code or baby name data can. So my thought is, are fancy visualizations a result of data that is conducive to producing them or can they be made for any dataset?

jessyue wrote:

I would also like to question Bertin's statement that decision-making cannot be automated. Today in the field of computer vision, image processing, machine learning, etc., decision-making gets automated by computers everyday, and this is still improving. For example, computer algorithms can draw some important conclusions on how to segment an image based on its histogram representation. Then, Bertin's definition of "extrinsic information" becomes vague because it is by definition, "everything else which cannot be processed by machine". The ability of computers to process information is changing at a rapid pace.

The example is interesting because the assistant separated each drawings on different cards. In a way, he/she made a non-computerized version of an animation by arranging the cards. Through this interaction of rearranging the information on each card, the assistant and manager was able to draw important conclusions. This example demonstrate the power of interaction.

jessyue wrote:

I would also like to question Bertin's statement that decision-making cannot be automated. Today in the field of computer vision, image processing, machine learning, etc., decision-making gets automated by computers everyday, and this is still improving. For example, computer algorithms can draw some important conclusions on how to segment an image based on its histogram representation. Then, Bertin's definition of "extrinsic information" becomes vague because it is by definition, "everything else which cannot be processed by machine". The ability of computers to process information is changing at a rapid pace.

The example is interesting because the assistant separated each drawings on different cards. In a way, he/she made a non-computerized version of an animation by arranging the cards. Through this interaction of rearranging the information on each card, the assistant and manager was able to draw important conclusions. This example demonstrate the power of interaction.

mlchu wrote:

I am impressed by the how the TimeSearcher makes use of TimeBox to extend the query expressiveness. I thought of TimeBox as a 2-D version of the slider bar we discussed in the lecture about exploring multidimensional data on parallel coordinates, and it enables user to make 2-way query on the dataset encoding time dimension. In my head I was thinking adding an extra slider bar on the time axis, together with the slider bar on the y-axis, would essentially yield the same query result but I am convinced that the use of a "box" is a better design for human graphic interaction, the same as we sometimes prefer point-and-select then typing.

One salient point that I learned after seeing different data query tools is that, the designer of the tool would have to make important design decision on the technique applicable to the unique dataset. For example, timebox works well with time series data, but perhaps not so for data encoded in a geographic map (like zipdecode); instead, slider bars for filtering quantitative variables or even text search are user-friendlier. I think that, with interactive tools, we are not only limited to convey message through static visualization. This opportunity, at the same time, presents us a challenge on considering the interaction experience and envisioning what the tool’s end-user want to query and explore.

jkeeshin wrote:

My favorite techniques discussed today in class that were discussed today that we had also discussed a little before hand were brushing and linking. Among the interaction techniques presented, I found that one to be the most novel. Some of the other techniques like pointing, selecting, direct manipulation, and data-on-demand seem more obvious. I think brushing and linking provides a really nice way to see a lot of data and also explore it at the same time.

I also found the hotel guest example from class and the reading to be very interesting. I think the way we reorganized the data, and then did width duplication to really highlight patterns was a really interesting technique.

The zip code and baby names examples were also a lot of fun--the baby names one because it introduced lots of surprising data in an easy to navigate way, and the zip code one because it provided a simple interface for showing data. It had a relatively simple data set, but the use of maps colors, and update as you type was really fantastic.

yeyleo wrote:

I enjoyed the zip code example as well as the baby names example that we saw in class today. However, one thing that stood out to me was that both data sets lend themselves very easily to an effective visualization. It could be because both have ways that they can be sorted as well as searchable, therefore making themselves conducive to fancy visualizations such as those that D3 provides. But how would one go about making a great interactive dataset for data such as SCL levels over a day? It doesn't seem like that type of data can lend itself to a "wow" effect the same way that the zip code or baby name data can. So my thought is, are fancy visualizations a result of data that is conducive to producing them or can they be made for any dataset?

pcish wrote:

It is always fun to read papers from before the turn of the century, such as the one on the cognitive coprocessor by Robertson et al. It took a while to realize that behind the fancy names terms The Multiple Agent Problem and The Animation Problem, what they are really describing (in the context of their "Information Visualizer") is really a virtual world representation that today is so common we hardly think about it. We need to look no further than the nearest video games to find a working example. System performance and people's expectations have moved far beyond the 10fps target of the work, and new interation techniques such as speech and motion have also appeared in the two decades that have gone by. Today's virtual reality technology is far more advanced than it was back in 1990, but still has a long way to go to be really fully immersive, and as this paper with its important-sounding terms made me realize, can also be a very good medium for information visualization.

vulcan wrote:

I found Bertin's "Postmortem of an Example" a helpful paradigm for thinking about how effective visualizations are constructed, and I liked his comment about how information in a visualization "can always be written or imagined in a single table of numbers." I thought back to Tableau and how easy it is to just select a bunch of dimensions and measures and click "Show Me" without really thinking about how the data is laid out in table form. Although Tableau's "Show Me" is smart, I feel like keeping in mind that visualizations are grounded in numerical/categorical data somewhere is important as well.

yangyh wrote:

Bertin's "Postmortem of an Example" is a interesting read. It was quite easy to understand, and it pointed out some critical point of views in decision-making. I found the five stages helpful in many situations, and I should definitely try to follow them. Indeed, in the modern society (especially during the high-tech boom like now), massive data sets are everywhere. How to effectively convey to people the "overall relationships" as Bertin suggested becomes an emerging field. As an engineer, it is definitely required to obtain some skills on large data analysis, as it helps so much if we know what data actually represent/what problem is actually lying down there. The stages Bertin suggested would serve as great guidelines to these sort of problem.

By the way, the baby name demo in class was awesome! I was surprised to see that seemingly tedious data set can lead to such an interesting visualization. It was great to see how data interact with people (with a lot of fun). I definitely would like to explore more in Assignment 3!

netj wrote:

Since the rise of GUI, I always had a doubt: is "selection by pointing" so superior to "typing" that avoiding it is a better design? It was interesting to see "selection by pointing (not typing)" among the key concepts of VIS reading and also in today's slide for Direct manipulation. I agree pointing is the most intuitive way to interact which became even more with the advent of touch screens, but there are still a lot of tasks where typing gives a better experience. We've seen good examples in the class: Zipdecode and NameVoyager. It'll be a nonsense to make users choose numbers from the screen to interact with Zipdecode instead of typing them, or forcing them to pick a tiny stripe of name from the stacked area chart to see them. By typing a few letters, users can instantly specify a pattern to shape the data they want to see. Maybe in those examples, due to the nature of data they are visualizing, typing might be the only way for direct manipulation?

The most apparent case where typing is superior to pointing is when there's a pattern language for interaction, e.g. glob pattern, or regular expression. Although typing can't beat point-and-click for selecting irregular targets, once they have a pattern, typing is much more efficient and accurate. Imagine you have thousands of files of data for several experiments, and you want to select those whose name looks like data-201110??-exp1.csv. Since your Finder or Explorer does not allow you to type them, you need to go through the huge list and select each of them.

Next one digresses a bit, but another traditional advantage of typing, especially on fixed devices like hardware keyboard, was that it could be done without even pointing! Our hands are usually positioned on it to simply press the correct key. Prior to the touch screens, most of the pointing devices were indirect manipulation which required synchronizing one movement on a different plane with the movement of cursor on the display, which creates a huge burden on the user. I believe this was why the scroll wheel came into existence, and why some people still prefer having hardware keyboard for their mobile phones, and many user likes to use keyboard shortcuts instead of searching through menus.

In short, I think to provide effective interaction experience for experts and novices, both typing queries in pattern languages and selection by pointing should be designed to work well together. Additionally, if the pointing devices are indirect, keyboard shortcuts and/or mapping of fixed hardware buttons should always be considered.

netj wrote:
bbunge wrote:

After reading Visual information seeking: Tight coupling of dynamic query filters with starfield displays, Ahlberg & Shneiderman and watching the demos in class today, I was curious to find out what the motivation was behind the Alphaslider. Why did this seem like a good idea?

The original paper is The Alphaslider: A Compact and Rapid Selector

Here is some of the rationale behind avoiding the keyboard: "Although selecting words or names with an Alphaslider might in some cases be slower than typing on a keyboard, the use of an Alphaslider has several advantages compared to a keyboard. Using a keyboard, inexperienced users must search the keyboard for the appropriate key and the keyboard does not prevent misspellings. Users may type a value for a field that is inappropriate such as a number when a person’s name is required [21]. An Alphaslider by definition contains all valid input choices and can continuously have its query range updated, which effectively eliminates queries that will result in an invalid or empty query result."

These cons are enough to warrant a confused facial expression. I suppose the audience at the time may have been different in terms of experience, but I can't imagine a case where someone would be likely to press the wrong key more than twice. Even a few times is much less than what I could imagine is made using a slider, especially for target letters that aren't inflated. It would have been interesting to see the results of typing as one of the conditions.

I do give them credit for a creative approach to working within their design constraints.

junjie87 wrote:

One thing I noticed about all the visualizations that I liked today was that each of them gave instant user feedback for every "unit" of interaction that the user made. I feel that the instant feedback made the interaction feel slick and smooth, and it also reduced the time it took for one to learn how to use the interaction confidently. Seen another way, instant feedback simultaneously reduces the gulf of execution (it must be easy to make changes) and the gulf of evaluation (if I made one action, and one thing changed on the screen, I can pretty much conclude my action caused the change).

About the Alphaslider, if the authors were worried that inexperienced typists might misspell names, they could have a textbox that gave suggestions as one typed. The suggestions can even match fuzzy names, so if I typed "jlo" it could match "Jennifer Lopez". Interestingly, this is what google did with google suggest and later, google instant, to make feedback appear as soon as possible, down to a single input letter.

schneibe wrote:

Interaction definitively adds something valuable to a visualization: the user can actively explore the dataset and is no longer a passive actor when looking at a graph. It made me think about the concept of epistemic action, that is "actions whose purpose is not to alter the world so as to advance physically toward some goal, but rather to alter the world so as to help make available information required as part of a problem solving routine. Examples of epistemic actions include looking at a chessboard from different angles or organizing the spatial layout of a hand of card" (Kirsh & Maglio 1994). Like chess players users can now take a look at the graph from different point of views, which decreases the cognitive load (due to mental computation). If you are interested in this concept, read the paper! It's extremely interesting, they studied this phenomenon by looking at novices and experts tetris players.

stojanik wrote:

@jneid, @schneibe, and @Bertin -

I enjoyed the common threads of your posts (Bertin exluded) with respect to the processes of interaction and problem solving (both automated and through human agency). Bertin's assumptions of what aspects of the "decision-making" process can be automated and how this relates to interaction is an exciting topic. And I find myself slightly agreeing with Bertin in that automation - machine intelligence - is not quite there to address the unique problem-solving capabilities of humans. Inselberg, in the Multidimensional Detective, also touches on this human desire for automation when he talks about "intelligent agents, gleaned from the accumulated experience" automating the exploration process. (p.107)

I think it's a great asset to have these machine-tools help us solve problems, but I am not sure what the rush is to completely displace human agency in this process. The context of problem-solving interaction is really important as I think that goes a long way to determining the results. A great example of collaborative human agency, coupled with machine-tools, triumphing over machine automation alone can be found here, Online gamers crack AIDS enzyme puzzle. The context of game-based collaborative problem-solving produced an accurate model of an enzyme in just three weeks that automation alone could not.

A few of the key points from the article:

We wanted to see if human intuition could succeed where automated methods had failed," Firas Khatib of the university's biochemistry lab said in a press release. "The ingenuity of game players is a formidable force that, if properly directed, can be used to solve a wide range of scientific problems.

One of Foldit's creators, Seth Cooper, explained why gamers had succeeded where computers had failed.
People have spatial reasoning skills, something computers are not yet good at

Games provide a framework for bringing together the strengths of computers and humans. The results in this week's paper show that gaming, science and computation can be combined to make advances that were not possible before.

awpharr wrote:

@abless: I think that you are right about the nature of static and dynamic visulizations. Using a static visualization results in only that information being interpreted by the viewer, so it is much easier for the viewer to come to a quick understanding of the data. In general, a dynamic display inherently requires a bit more exploration for it to be fully effective. As @zgalant mentioned, this dynamic nature makes the NameVoyager visualization much more appealing and informative than the static CS106A equivalent. Therefore, I believe that when deciding to use a static or dynamic visualization, we should really be looking at what the logical process of the viewer looks like. If they are going to be using it more for exploratory purposes, a dynamic visualization, when applicable, is probably going to come across more strongly than a static one. When the data being visualized is really only telling a few distinct stories, a static visualization is probably going make these stories much clearer and easier for the viewer to quickly discern.

tpurtell wrote:

I was looking up some more information on seriation and I found a paper about matrix visualization. There was another interesting type of plot they showed called a sediment display. Each column is sorted independent of the row. This lets you see a distribution for each of the different rows in the data. Many of the visualization in this text use color as a means to encode a quantitative variable. Although it is not the most effective for giving a strong impression of a specific value, it seems to do a reasonable job letting the viewer see relative relationships between data.

Another example of a graphical matrix is presents views of different quantities expressed over the matrix data. These plots are really dense, but I suppose they can be quite helpful in exploratory analysis of certain types of data.

How well do these type of techniques hold up under time series animation?

kpoppen wrote:

@jneid although I appreciate your arguments about automating the decision-making process, I feel like automated outlier-detection and consequent decision-making is still very much constrained (at least currently) by the imaginative process that Bertin refers to is still ultimately done by humans. The only difference is that rather than humans coming up with decisions directly, they instead reason about what kinds of decisions you would want to make given certain kinds / magnitudes of outliers. I don't feel like the end to end system of "generally observe a whole bunch of datapoints to detect outliers, and then automatically identify actionable attributes of these outliers so as to make a decision" is something, given the fickle (and/or specific) nature of modern machine learning. This is not to say that such a system would not be possible so much as it is to say that I think Bertin's real point was more that ultimately humans are responsible for the creative task of qualifying the initial problem, and what decisions can be made, independent of the autonomy of computers in carrying out the process thereafter.

mkanne wrote:

@stojanik - I really enjoyed the article about Foldit you linked to and believe that this human achievement really highlights that problems can be solved most efficiently by taking advantage of the unique computational and reasoning skills of humans and computers according to what skills are required. Most of us can not look at a table of data and visualize the relationships in our minds but a computer can easily transform that data to reveal relationships. Similarly, computers have a much harder time with spatial reasoning (as the article states) while we can look at a visualization and pre-attentively pick out patterns with little cognitive load.

Bertin's enumeration of intrinsic vs. extrinsic factors reflects upon these ideas. Others have argued above that Bertin saw a dichotomy where there is none because he did not have foreknowledge of the current capabilities of computers. I believe that such a dichotomy does still exist although not exactly as described by Bertin. I was also intrigued by Bertin's argument that other data that can't be put in the table ("number of rooms, beds, people, repairs per floor" page 3) is not useful in the visualization. In fact, comparing this data with the table data may reveal important trends. It would also yield a non-time series visualization that would be "better" by Tufte's standards.

I also really enjoyed the timeboxes article and loved that idea of dynamically highlighting data sets based on user provided parameters. What made the timebox solution great, in my opinion, was its well thought out usability. By providing users with an interface that uses familiar data selection techniques, timeboxes made data exploration (as stated in the article) cognitively simple.

ashtona wrote:

This lecture was super interesting. The problem of sticking to dataviz fundamentals Tufte-style seems doubly hard when you throw interaction into the mix. For example, the demo Jeff showed with the stock charts where you could choose boxes that the stock trend lines had to fall inside to be displayed was really cool, but I didn't find the method of selecting the error bounds on the derivative that intuitive (you adjusted the slope of a line to match the derivative, and then the length of a vertical line denoted your error bounds). Kind of haphazard. Anyway, that was a small quibble -- I'm just pointing out that it seems way easier to make a fundamentally flawed design decision with interaction.

Also never noticed how powerful simple selection could be. The "linking" technique is so simple yet such a great way to quickly visualize correlations across multiple dimensions.

blouie wrote:

I likewise found Bertin's reading to be quite interesting. In general, I think the process flow makes a lot of sense, especially given the part about reclassing data to let it communicate something useful. But the reading reinforces the point that data visualizations are powerful when they are able to provide insight on a particular argument. So in the final step of the process flow, I feel like Bertin didn't take it far enough. I feel like he needed to say "From this data, we conclude _," which would guide the user into looking for a certain pattern or conclusion when looking at the data. True, data visualization minus this extra step might be good as an exploratory tool. But I wonder how often that is: I wonder exactly how often data is presented to another person (i.e., the manager) without there being some kind of slant that the presenter hopes the observer takes away from the visualization.

zhenghao wrote:

The lecture today on interaction really hammered home how effective some simple user interactions like selection and animations can be. I also feel that it seems fairly under-appreciated that beneath all the discussion about effectiveness in conveying information or facilitating certain cognitive processes that interactions also make visualizations fun. I must confess that I've spent more time than I should playing around with some of the visualizations around on the internet just because I like watching the transitions or because they're strangely addictive : P

ashtona wrote:

This lecture was super interesting. The problem of sticking to dataviz fundamentals Tufte-style seems doubly hard when you throw interaction into the mix. For example, the demo Jeff showed with the stock charts where you could choose boxes that the stock trend lines had to fall inside to be displayed was really cool, but I didn't find the method of selecting the error bounds on the derivative that intuitive (you adjusted the slope of a line to match the derivative, and then the length of a vertical line denoted your error bounds). Kind of haphazard. Anyway, that was a small quibble -- I'm just pointing out that it seems way easier to make a fundamentally flawed design decision with interaction.

Also never noticed how powerful simple selection could be. The "linking" technique is so simple yet such a great way to quickly visualize correlations across multiple dimensions.

ashtona wrote:

This lecture was super interesting. The problem of sticking to dataviz fundamentals Tufte-style seems doubly hard when you throw interaction into the mix. For example, the Hochheiser & Shneiderman time series visualization where you can choose boxes that the trend lines had to fall inside to be displayed is really cool, but I don't find the method of selecting the error bounds on the derivative that intuitive (you adjust the slope of a line to match the derivative, and then the length of a vertical line denotes your error bounds). Kind of haphazard. Anyway, that was a small quibble -- I'm just pointing out that it seems way easier to make a fundamentally flawed design decision with interaction.

Also never noticed how powerful simple selection could be. The "linking" technique is so simple yet such a great way to quickly visualize correlations across multiple dimensions.

fcai10 wrote:

Like the other posters, I really enjoyed the interactive examples in class. I felt the examples, especially NameVoyager, really engaged the class and that reminded me of the idea from one of the previous readings where the author suggested that you can judge how successful your visualization is by how much you start playing around with it.

Echoing the sentiments of those who posted before me, I also enjoyed the readings for giving names to ideas I take for granted -- such as tight coupling, output-as-input (Alhberg and Shneiderman). In general, I've come to realize how powerful it is to be able to *name* something -- it will definitely make me more attentive to the design choices that go into visualizations I see and that I will design. I also found the quote at the end of the Bertin reading interesting: "Information is the reply to a question", although I'd like to suggest that information may contain the answer to a question, but is not necessary the answer itself. Semantics aside, I think the more general direction of the quote is to suggest that when we are visualization information, we should have a question in mind -- such as in A2, when we were encouraged to have a driving question as we explored our data.

bsee wrote:

This lecture made me realize that I have been taking dynamic query filters for granted my entire life. I didn't even know that it had a name. I don't know how it would be like to find air tickets without dynamic query filters... ... :P

I would like to touch on the point raised during lecture about efficiency and expressiveness. This point is really similar to the point raised by someone on Tuesday, on if we can make D3 both intuitive and flexible. Apparently this slide just reiterated the fact that they are probably on polar ends of the scale. I guess it ultimately boils down to the question, is this tool "good enough"? If we just want to convey an increasing or decreasing trend, then it makes no sense to use complex api to do that. Similarly, if we need to include interaction, then trying to use Excel to do that is not feasible either. Thus, knowing the scale and complexity of the problem is crucial in design process.

rc8138 wrote:

This lecture really widen my view on data visualization. As a statistics student, emphasis are often put on the mathematical derivation. The aim, of course, is to infer and draw sensible conclusion from the data. Visualizations have been important ways for statistician to communicate with others, but in traditional curriculum, interactive displays are rarely used.

I especially appreciate how we can view the different visualization tools from the perspective of efficiency and expressiveness. My (perhaps biased) opinion is that the general statistics community tend to sacrifice expressiveness for efficiency because the math is the main training. I am very happy to be introduced to these new possibilities, and will certainly keep in mind the two metrics when deciding what tool to build my visualization.

The concept of "Brushing and Linking" and "dynamic querying" seem very natural, and both techniques really motivate the users to "interact" with data. This is very different from static visualizations, where users' main task is to decode the designers encoding of information. In interactive visualization, users find useful information on their own based on the design of the visualization.

jsadler wrote:

One of the things Jeff called out in the lecture was the "ribbon" indexing feature show in filmfinder.

With this interface the designers seemed to want you to drag a slider across A-Z in order to select the first letter of the movie etc.

This reminded me of another place I have seen this interaction - the iphone address book...

Interestingly in the first iOS version of the iphone the only way you could select a contact was by touching and dragging through such a A-Z ribbon. Compare this with typing in the first two letters of the persons name (that ability to search by typing the persons name was introduced later.. thankfully)

It was interesting to see this interaction resurface in the case of iphone, and feel the consequences ...

/wikis/cs448b-11-fall/Interaction/Discussion-2011-10-23-22-06-18?action=AttachFile&do=get&target=iphone.png

dsmith2 wrote:

When showing us some examples for Assignment 3 of interactive visualizations (baby names, zip codes), I was particularly struck by one quality of the data: simplicity.

Bertin comments that simplicity in visualization is critical so that our eyes simplify that which we see. Many of the above posters have commented on simplicity of the interactions and visualizations and how powerful visualizations should use simplicity to avoid the need for explanation.

A few people mentioned data simplicity in passing, but it seems to be overshadowed by the discussion of visual simplicity. While this makes sense due to the course material, I have found from assignments 1 & 2 that when picking my own data sets, I try to look for interesting nuanced relationships and overcomplicate the search for data.

The zipcode finder was a refreshing and inspiring example of how a truly successful visualization can take a rather ordinary collection of information that one is used to seeing and reframe it in a new context to lead to a new understanding.

Finding the right data is possibly the most daunting task of creating visualizations. I have found it easy to get bogged down in mounds of information, often chasing that which is difficult to collect, with the assumption that it is more valuable. I hope the zipcode and baby name examples allow me to find simple but captivating data sets that I hadn't before understood to be interesting.

jhlau wrote:

I think the timebox mechanism in the Hochheiser Schneiderman and demonstrated in class is a great interaction technique. I think it especially excels in data that can be too complicated to visualize easily in an effective manner, such as data sets that include a large number of time series data (it becomes difficult to graph this data without making it look like a giant ink splotch). I also find that this technique becomes very powerful when combined with small multiples (such as in one of the examples in class). The combination allows timeboxing on several attributes, which I feel is essential to be able to effectively study anything but the simplest data sets.

But speaking of simple data sets, it amazes me how the best visualizations seem to come from simple, rather than complex, data sets. In my opinion, the most effective visualizations show in class were the baby name explorer and the zip code visualization. Both data sets were extremely simple, having only two or three variables, but the visualizations were extremely informative. The question I have: do simple data sets tend to create more effective visualizations, or were these two data sets just extremely well-chosen? In other words, is it more about choosing a simple data set, or choosing a specific data set with certain characteristics?

Finally, I'm thinking back to the ink-data ratio principle from earlier in the quarter. It seemed to me that the intuitive analog for interactive visualizations would be interaction-data ratio, but after reflecting on the visualizations shown in class, I don't think this is the case. For example, while the baby name explorer could've used static refresh transitions between visualizations, it seemed that adding transitions was actually a very effective way to link the previous visualization with the current visualization (say, when you were drilling down on a subset of names). I'm curious to see what I think about this topic after seeing more visualizations.

insunj wrote:

I found both VIS and Visual exploration of time-series data to be enjoyable and informative. Time series paper covered exploration of data using constraint box at different times, which was also covered in the lecture. I found that although this paper tells you what they have done, but doesn't prove or show you why it is a good technique. It was nice to learn about applications to this technique at the end though. VIS paper, on the other hand, covered the basic principles and rules for the interactive visual designs.As we talked about in the lecture, there is definitely ways to improve dynamic queries like the film finder. I wonder for dynamic queries, what constitutes a good set of interactive queries, given enormous data! what when the categories big, what are the ways to represent them nicely other than colors or shapes?

sakshia wrote:

It was interesting to read the 3 papers together and see how each addresses a slightly different user. The Post-Mortem paper and the TimeSearcher could be considered more similar - because they explore data sets which are time series, and are from the perspective of someone who is trying to query to find trends, i.e. a user investigating the data. On the other hand, the example explored in the VIS paper, FilmFinder, focuses more on a user who is a 'consumer' and is looking to make a choice out of a large set of data. To be more precise, the users have different goals - the former looks more at discovering trends and classifying information, whereas the latter is more about search. I wonder how the principles suggested in the VIS paper apply to the TimeSearcher and hotel manager examples.

babchick wrote:

I'm a big fan of the generalized selection paper and the effective use of direct manipulation used to achieve it. One of the more interesting underlying principles of direct manipulation was the "output-is-input" aspect of tight coupling. The general ability to take the solution to an information need and use that result to expose more opportunities for exploration is something extremely powerful, and this connectedness via tight coupling is something that I think people secretly cannot get enough of. Facebook is a great example of putting this output-is-input principle to addictive use, where almost every single bit of information on the page is becoming clickable to "dive down" into that particular object/person/event page on the open graph.

There has been a lot of discussion about simplicity and boiling down a visualization to its essentials, and I would like to suggest a more concrete heuristic for "essentials" for an interactive technique as some data and presentation that allows for this 'output-is-input' style of tight coupling; it seemed to lie at the core of what made the Interactive Query Relaxation paper so compelling.

dbrody wrote:

From the required readings it is apparent how important interaction for data visualization is. While its really interesting to see how the interaction has changed over time, from files and papers to mouse and touch, the main purpose continues to be important. The ways Berg and Scott presented to be able to manipulate and repurpose the data to easily see trends is really interesting. They definitely took a lot of data and structured it in a way that makes it a lot more feasible to find trends and to manage. Furthermore, their use of different design techniques of color coding and positioning enhances the way the user can interact and explore the data. Their points they highlight are really crutial I think in presenting large amounts of data and will definitely be useful in designing visualizations in the future.

Leave a comment