Skip to main content
cancel
Showing results for 
Search instead for 
Did you mean: 
stephencredmond
Luminary Alumni
Luminary Alumni

A Challenge - just for fun

See http://community.qlik.com/media/p/70295.aspx for the QVW

In the following BBC article: http://news.bbc.co.uk/2/hi/uk_news/magazine/7937382.stm, the chart below was touted as being an excellent way of visualising data. I think that it is crap.

Propose a better way using QlikView objects.

Stephen

31 Replies
johnw
Champion III
Champion III

I don't think we've hit the QlikView charting wall. I think we've hit the SINGLE STATIC QlikView charting wall. But QlikView is an interactive tool for data analysis, not a single static chart that displays everything you might ever want to know all at once.

In other words, asking us to create a single chart that shows as much information more clearly than the original is to not take advantage of QlikView's strengths.

If this were a real application, I'd probably have a dashboard with two or three simple charts to make the big patterns jump out. I'm fairly confident they could be presented more clearly than the original chart, which fits "too much" information in, and as a result seems confusing at first and requires study to find the information. So break it apart and use a consistent theme like bar graphs with consistent and meaningful colors. Then allow for selections and cyclic or hierarchical groups and so on to dig in further. Probably include a simple pivot table with counts and percentages for when you want to know the hard data. I think it would work quite well.

Certainly, it would be nice if QlikView were even better - if had more types of charts including this one and stronger, simpler support for user customization of charts (I agree with you that user objects are too complicated to count). But I think for most users, you want to present the information as simply as possible, with as basic a chart as possible, and as few distracting elements as possible. Everyone understands a basic bar chart. And at least my users don't seem to want to play with interactive charts beyond just making selections and possibly a cyclic dimension here and there. I've gone to quite a bit of development effort to give them the ability to add and remove dimensions and expressions, change the chart types, and do things like that. I'm fairly confident none of them use it on any sort of regular basis, no matter how powerful it is and how well it would answer their questions. So while it would be nice for QlikView to more natively support what I'm doing, I don't think it's actually that important most of the time.

Not applicable

Well said, John.

Changing ideas here, but I'd be much more enthusiastic about the original chart if, say, it were possible to swap the horizontal position of Male/Female. And if I could hover over one of the colors and see how the numbers changed moving vertically.

stephencredmond
Luminary Alumni
Luminary Alumni
Author


From: Stephen Redmond
Sent: Thu 16/07/2009 00:16
To: Stephen Few
Subject: RE: New chart type

Hi Steve,
Thanks for the links. I am still not totally enamoured with Kosara et al in this visualisation. I think that I can see that they are trying to build a very flexibile tool, but I feel that they are losing something.
I first heard you speaking at the QlikTech conference in Miami last year. After > 10 years working with various report tools (mostly Crystal) and now > 3 years on QlikView, I was very taken with your ideas. I have given up the pies - though my waistband hasn't felt the benefit yet. I strive for a balance between good visuals and ease of use and I like to think that I have taken on board the ideas of your first two books (I am waiting for a colleague to surrender the new one).
I have read your 2 articles on Visual Multivariate Analysis and the subsequent one on Parallel Coordinates. All very good (although not a fan of the parabox). I just haven't been sold on the use of the Parallel Sets in this way. The fact that a user would manipulate the chart to achieve a better shape means, to me, that there is no intrinsic shape that this chart will produce out of the data and different people will produce different charts from the same data and come up with different conclusions.
All subjective of course.
Would I be totally wrong to propose that a better visualisation for the "Titanic" data would be the humble pivot table?
Thanks for your excellent work in this area. I look forward to your next article.
Regards,
Stephen
stephencredmond
Luminary Alumni
Luminary Alumni
Author


From: Stephen Few
Sent: Thu 16/07/2009 17:28
To: Stephen Redmond
Subject: Re: New chart type

Hi Stephen,

The problem with the pivot table, and all tables, is that we can only see relationships among the various groups by constructing them piece by piece from individual values in the table, trying to hold them in our heads as we do so, which is impossible due to the limits of working memory. To see and examine these relationships, we need a picture. In the case of the Titanic data, the picture must display relationships among three categorical variables--sex, class, and survival status--with counts of people who fall into each intersection between them. It must do so in a way that doesn't just make it easy to look up individual values, such as the number of female crew members who survived, which the table does well, but in a way that let's us easily see and compare all of the groups at various hierarchical levels (all people, all survivors, all females, all 1st class passengers, first class passengers that survived, female passengers that survived, female 1st class passengers who survived, etc.) at the same time. Kosara's parallel sets is an attempt to visualize data sets like this as well as those that are much more complex in that they involve many more variables. Rather than three categorical dimensions, imagine 10. I suspect that we could come up with a better way of visualizing the Titanic data than by using parallel sets (for example, a visual crosstab of bar graphs, such as those that are easily constructed using Tableau, and with a bit more work using QlikView), but whatever we come up with would probably cease to be effective if more variables were involved.

I really appreciate the fact that you're thinking about these things, exploring various approaches, and trying to come up with the best solutions. This is design thinking, which much too rare.

Take care,

Steve

johnw
Champion III
Champion III

I mostly agree with Stephen Few on this one, but not entirely. For instance:

"The problem with the pivot table, and all tables, is that we can only see relationships among the various groups by constructing them piece by piece from individual values in the table, trying to hold them in our heads as we do so, which is impossible due to the limits of working memory."

With the small number of variables here, and the subtotals available, we generally only need to look at a couple of values to answer a question, and can answer questions at more than one point in the hierarchy. So it doesn't seem as weak as he makes it out to be for this specific data set, even if I agree that he's correct in the general case.

Anyway, I want to look at this statement of his more closely:

"It must do so in a way that doesn't just make it easy to look up individual values, such as the number of female crew members who survived, which the table does well, but in a way that let's us easily see and compare all of the groups at various hierarchical levels (all people, all survivors, all females, all 1st class passengers, first class passengers that survived, female passengers that survived, female 1st class passengers who survived, etc.) at the same time." (emphasis added)

What I DON'T think is necessary in QlikView is having a single chart that can answer all of these questions at the same time. Rather, we're good as long as we can easily answer all of these questions. Multiple charts are fine. One or two clicks is fine. So a QlikView pivot table, for instance, can answer ALL of these questions by moving the dimensions around, just not all questions at the same time. Or you could have multiple charts that do it. Or some combination of the two like I proposed so that you can see the "most important" facts at a glance, and look up any additional information with just a few clicks.

His mentioning of "hierarchical levels" is perhaps a better way to say what I was trying to say with "summary information". The original chart allows data lookups at all hierarchical levels all in the same chart, and all visually rather than with numbers. In comparison, most of the solutions we've presented in this thread are only at a single level in the hierarchy. But I also like to stay away from "hierarchy" as there is no single hierarchy for this information. There are multiple possible hierarchies. And frankly, I don't think of it as a hierarchy at all. I see three dimensions. Actually, I see four. A person (could just be recno() here), sex, class, and whether they lived or died. No hierarchy implied by that in my opinion.

Interestingly, he mentions a visual crosstab of bar graphs, which sounds like the trellis chart solution by kjn, which does suffer from only being at a single level in the hierarchy. Perhaps he means something different than that. I'm not up on the lingo.

I should probably read the articles involved.

rwunderlich
Partner Ambassador/MVP
Partner Ambassador/MVP


Jay Jakosky wrote:Beginning with a bar chart, the user should be able to interact with it how they want. A pivot table allows the user to swap column order or collapse lower-level aggregation, but charts do not.


How about fast change? Switch to a pivot, move the columns around and switch back to chart.

-Rob

Not applicable


Rob Wunderlich wrote:
<pre>
Jay Jakosky wrote:Beginning with a bar chart, the user should be able to interact with it how they want. A pivot table allows the user to swap column order or collapse lower-level aggregation, but charts do not.


How about fast change? Switch to a pivot, move the columns around and switch back to chart.

-Rob

Indeed, for that situation, but what is the larger end-user experience? We all share the experiences of extracting, transforming and exploring data with QlikView. By comparison to any other BI tool, it's a pleasure to use.

Download Tableau and you will have that kind of experience for visualization. By comparison to any other tool, including QlikView, it's a pleasure to use. Its treatment of geographic data, scatter plots and trellis charts is smooth and attractive. The automatic chart selection is very good. It is dead simple to add, remove, pivot, swap and combine dimensions and metrics.

Users have the same "go ahead and explore" experience with visualization in Tableau that QlikView users have with data. I think QlikView should have the same quality in its presentation.

rwunderlich
Partner Ambassador/MVP
Partner Ambassador/MVP


John Witherspoon wrote:Reasonable at displaying some of the facts, and probably better than the multiple pie charts. But what percentage of passengers were female? What percentage of male passengers survived? What percentage of female passengers were in first class? Your bar chart mostly gives you information at the lowest level of detail. Summary information requires you to mentally add together the lengths of bars or bar segments. In the original, these summary facts are easily seen once you figure out how to read the chart.


I'll agree that my chart does not provide as much "data", but I'll argue that it presents more "information".

I don't think exact and total percentages are important or useful here. I was inspired by Jay's first submission and his comment "It was good to be a rich female on the Titanic". That led me to approach this more as a graphic than an analytical chart. Edward Tufte says to "know what story you want to tell before you design the graphic".

Stephen's challenge was a "better way to visualize the data". I think my submission qualifies in that it brings out the story more clearly than the original chart. A lot of people perished, more than survived. The two color pattern makes this immeadiate. The mostly male Crew gave up a lot of souls. There is an ordered economic heirarchy which aligns with loss. The upper class women fared extremely well. etc etc. I can hear the creak of the hull.

-Rob

(BTW, I pulled my numbers from Stephen's posted example. The published numbers I've found are much greater, but the proportions are similar)

johnw
Champion III
Champion III


Rob Wunderlich wrote:I don't think exact and total percentages are important or useful here.
I see your point. Your graph, with the colors and with the widespread understanding of bar graphs, has much more immediate impact, and communicates some of the important points very quickly and intuitively.

First thing is that there's a lot of red. The human mind will pick up on that immediately. You don't even have to read the legend first, since most people will assume from context that red is dead. In the first fraction of a second, you'll have a decent comprehension of the magnitude of the deaths. That understanding will be hard to quantify, but perhaps knowing "a lot more than half" in a fraction of a second is more important than knowing "about 2/3" ten seconds later, when you've figured out how to read the original graph.

After that, we can start looking for a general understanding of other trends. Even though we don't have totals or percentages or line lengths to easily compare, we can still see some of the basic concepts quite quickly. Women had a much better survival rate than men. How much better? We can't tell, but we have a basic feel for it. If backed up by an actual table of data, that feel would be plenty. First class passengers faired a little better than second class passengers, who faired much better than third class passengers, who faired similarly to the crew. Again, we can't tell much about the specifics, but we have a decent feel for it.

So perhaps given that understanding, a good QlikView solution would just be two charts. First chart would be the one that you presented, but with all of the legends removed (replace "1st" with "1st Class" and so on, and the Yes/No flag with Lived/Died). The second would be some sort of dynamic pivot table with numbers and percentages that could be fiddled with to produce any number you might want to look up after getting the feel for it from the original bar chart. Or if we aren't allowed dynamic solutions, some big lookup table that included numbers at every level of detail.

And while I'm switching sides on this, I might as well point out what I consider to be a weakness in the original chart - the human mind will assign significance to the AREAS of the parallelograms in the chart. But since the angles on these are arbitrary, the areas are arbitrary, so you might be fooled. Fortunately, the human mind isn't any good at evaluating relative areas, and the angles aren't particularly steep, so it won't interfere much with your understanding. But I still consider it a weakness, a misleading bit of non-information in the chart.

stephencredmond
Luminary Alumni
Luminary Alumni
Author

Hi All,

I am going to vote Rob's response as the winner. (Unfortunately, I appear to have lost my ability to mark the response as the "answer")

Certainly from 30,000ft it gives as much information as any single QlikView chart could.

I have found this thread very interesting and educational. It has been great for me.

Now, if you could create a chart to represent that data - any chart that you can imagine, whether it exists or not - what would that look like?

Stephen