I've heard people say that plots produced by ORIGIN tend to look polished and "professional," whereas plots produced by Mathematica do not. However, most plot-creation programs are quite configurable and it stands to reason that with the right settings for things like tick location and labeling, font and color choices, label alignment, and so on, I should be able to make a figure with Mathematica/matplotlib/Gnuplot/etc. that looks as good as those that come from ORIGIN. But what does it mean for a figure to be "professional" in this context?

In other words, if my goal is to create the best looking figures possible for inclusion in a scientific paper, what design choices are generally recommended towards that goal? Obviously one has to choose the appropriate kind of plot, e.g. bar graph vs. scatter plot, and linear vs. logarithmic scale, but those are choices that we always think about regardless of which plotting program we are using. I'm more interested in the things we don't normally think about, which are normally set according to some plotting program's defaults, but which could be changed to improve the look of the plot.

  • 2
    $\begingroup$ This may well be off topic here... I'm posting it to probe the boundaries of what acceptable data visualization questions are (c.f. meta.scicomp.stackexchange.com/questions/55/…) $\endgroup$
    – David Z
    Dec 12 '11 at 1:24
  • 5
    $\begingroup$ Considering the poor quality of many plots that make it into scientific publications, it may be that the community's definition needs improvement! $\endgroup$ Dec 12 '11 at 5:53
  • 2
    $\begingroup$ @DavidZaslavsky - I think this might actually be the most appropriate place for this question as it is specific to the presentation of scientific data, which is part of computational science. I have often seen scientists complaints that the graphs produced by our software are not up to 'publication standard' and have to be touched up, so it it would be nice to see other peoples experiences with this problem. As always though with subjective questions, answers should follow the six guidelines. $\endgroup$
    – Mark Booth
    Dec 12 '11 at 12:22
  • $\begingroup$ Of course, @Mark. I've tried to make this as objective as possible given that it's not a technical question. $\endgroup$
    – David Z
    Dec 12 '11 at 23:52
  • $\begingroup$ Thanks David, my comment was more aimed at those answering questions, as we were tending towards short answers at the time. The six guidelines prefer long over short, experience over opinion, context over assumption, impartiality over prejudice, serious over flippant and suggest backing up answers with facts and references etc. All these seem to be good things to strive for in answers to this type of question. $\endgroup$
    – Mark Booth
    Dec 13 '11 at 0:12

IMO, what makes a figure "professional quality" is defined by the journal/publisher rules. Which actually translates to "publication quality" which is relative depending on where you publish. Some universal rules seem to stand out - invariant of the plotting software being used:

1) A figure should contain as few elements possible required to convey the information/idea/argument. A figure should be easily read/understood within few seconds - if it takes you longer to understand what is going on on the figure, it might be too much information. This is sometimes hard to check as you are familiar with your own plot/data - no matter how over-encumbered it is - showing it to few colleagues to see if they can read it easily helps. (not to be mistaken with understanding the physical meaning behind the figure - this usually takes quite some time).

2) If you have to use colors, it is best to limit yourself to few, ideally on the opposite side of the color wheel. E.g. blue and red is better than blue and green. A figure may have many shades - but it is best to have few main colors. I often like to use blue (low values) and red (high values) with a white transition in-between. Always have in mind color-blind readers.

3) Tick marks, contour labels etc. all should be easily readable without a magnifying glass - so, similar font size as the journal body text. You can check if everything is readable by printing out a hardcopy with figure widths of 3 and 6 inches (these are common figure sizes in scientific journals).

4) Finally, make sure every single element of the figure has its purpose. If there is anything that is not conveying useful information - throw it out. It will help the readability of the figure.

At the point where you are comfortable with customizing all the little elements that make a figure - tick marks, labels etc., it does not really matter which tool you use as long as you are able to produce a clean eps.

  • $\begingroup$ Can you elaborate on the rationale for (1)? A few seconds is not a long time. Is this on the pragmatic grounds that people are really, really, busy, so that if it takes more than a few seconds to understand the figure, they are not going to bother? $\endgroup$ Dec 13 '11 at 19:59
  • $\begingroup$ @FaheemMitha No, what I really meant is, you are doing readers a favor if you keep the figure clear and simple. Don't hold on to few seconds as a solid criterion - this is just how I am used to evaluate my own figures. There is a limit to how much information you can put on one figure - by putting too much, you may end up making the contents hard for the reader to understand. Your readers will bother reading even very busy figures if they are interested in your work - but a less busy figure does a better job of conveying the information. $\endgroup$ Dec 13 '11 at 20:24
  • $\begingroup$ I'm not sure it entirely makes sense to accept an answer to such a nebulous question, but this answer seems to best codify the vague ideas I had in mind when I asked it. So you get the checkmark :-) $\endgroup$
    – David Z
    Dec 19 '11 at 8:24

There are a couple elements I look for when I consider something "publication-quality" in either my own work, or what I'm considering when looking at others. They are:

  1. High resolution, and preferably vector-based. This one should be fairly obvious by now, but you'd be surprised.
  2. A lack of clutter. I should be able to see what's happening in your figure, and see it quickly. There's few things I hate more than someone trying to take the "High Ink:Paper ratio" guidance and using it to try to cram an entire manuscript in a single figure.
  3. Prints well. This is the one that's actually most important for me, and when I'm reviewing papers, one I always test. "Do the figures print?" More than once, I've hit figures whose points are completely obfuscated when printed in grayscale, which renders them worthless for my purposes (I don't read on screens).
  4. Evidence that the creator knows how to use graphics settings. No odd-ball axis choices, tick marks in the right place, etc.
  5. Combined with #2, a lack of "flourish" that's entirely graphical in nature. Shadows, needless 3-D, etc. that really do nothing but waste the readers time.

Most of those are honestly creator-specific, rather than program specific. I've seen terrible plots done in R, and excellent plots done in Excel.

  • $\begingroup$ Vector graphics is indeed desirable. But are there any widely used drawing programs that are not? $\endgroup$ Dec 13 '11 at 3:12
  • 2
    $\begingroup$ Drawing programs? Not that I can think of. Programs that produce plots - statistical packages and the like? Many of them have non-vector output formats either available, or as their defaults. $\endgroup$
    – Fomite
    Dec 13 '11 at 3:43

If we're talking about data figures I'd go to the sources: Edward Tufte's The Visual Display of Qualitative Information and Beautiful Evidence.

Mr. Tufte of course goes into some details, but the principle that stands out for me is not spending ink on frames and decoration, but instead making as much of your ink as possible carry information.

Amended per Mark's request:

Some major points from The Visual Display of Qualitative Information are

  • show the data in a way that does not distort or obfuscate what it has to say
  • arrange displays to allow for comparisons between different data at different levels
  • integrate the graphic aspects with the statistical and verbal descriptions
  • maximize data to ink ratio by removing elements that do not serve any purpose (or are made redundant by other elements) and use what elements there are to convey additional information (axes that are variants on a boxplot, for example)
  • small multiples can be used to arrange higher dimensional data sets to allow comparison along these additional dimensions

Beautiful Evidence is a wider ranging book in its scope. I'll just reproduce the chapter titles:

  • Mapped Pictures: Images as Evidence and Explanation
  • Sparklines: Intense, Simple, Word-Sized Graphics
  • Links and Causal Arrows: Ambiguity in Action
  • Words, Numbers, Images -- Together
  • The Fundamental Principles of Analytical Design
  • Corruption in Evidence Presentations: Effects Without Causes, Cherry Picking, Overreaching, Chartjunk, and the Rage to Conclude
  • The Cognitive Style of PowerPoint: Pitching Out Corrupts Within
  • Sculptural Pedestals: Meaning, Practice, Depedestalization
  • Landscape Sculptures

One of the interesting observations in Beautiful Evidence is that we generally use high density output devices (a 300 DPI printer is a low density device these days) for printed material, but often draw our figures for screen or line printer, which wastes an enormous potential for conveying information.

  • $\begingroup$ Page 13 of The Visual Display of Quantitative Information starts off with a nice, bullet point, summary of "professional graphics," or as it is titled there "Graphical Excellence." $\endgroup$ Dec 12 '11 at 21:37
  • $\begingroup$ @Brian Diggs - Could either you or dmckee update this answer with a summary of the main points? Many of us may be interested enough to know the summary without being interested enough to actually goi out and buy the books. $\endgroup$
    – Mark Booth
    Dec 13 '11 at 0:06
  • $\begingroup$ @Mark: I'll get to it in the next day or so...my copy's at the office and right now I'm not. $\endgroup$ Dec 13 '11 at 0:29

The best figures I have been able to make personally have been with the TeX package PGF/TikZ. If you use LaTex, as many in the hard sciences do, you have probably already heard of it.

It also appears to be the leader in LaTex graphics packages. A sizeable proportion on the questions on the TeX StackExchange site are about PGF/TikZ. I'm not sure why the results are so good, but certainly one advantage PGF/TikZ has over other packages when using LaTeX is that it simply integrates better with the text. For one thing, the fonts in the figure will be the same as in the text.

  • $\begingroup$ Of course, actually it's my favorite way to make figures as well ;-) But I am more interested in why TikZ plots are considered good. $\endgroup$
    – David Z
    Dec 12 '11 at 1:37
  • 1
    $\begingroup$ @DavidZaslavsky: I dunno. Till is a genius? :-) $\endgroup$ Dec 12 '11 at 1:39
  • $\begingroup$ @DavidZaslavsky: Seriously, that question is bang on topic for tex.sx, and likely has already been addressed there. $\endgroup$ Dec 12 '11 at 1:41
  • 1
    $\begingroup$ @EpiGrad: I just meant that asking specifically about TikZ would be more useful on tex.sx, if David was so minded. Of course the broader question would not be on topic for tex.sx. $\endgroup$ Dec 13 '11 at 19:52
  • 1
    $\begingroup$ pgfplots is pretty amazing. By default it looks pretty much how you want it to, i.e. boxed, vector-crisp, labels consistent with body text, et. But it's also easy to adjust. My favorite trick is stripping the axes off of a Matlab pseudo colour plot, cropping it, and then wrapping axes around it in TikZ. That way you get a raster image for the pseudo colour (this is one of the few cases for which a raster graphic is preferable) and LaTeX goodness for axes and labels, and even the colour bar. This is one of my answers using pgfplots $\endgroup$
    – qubyte
    Dec 14 '11 at 5:17

It's almost easier to characterize what constitutes a bad graph than what makes a graph good.

Some features of bad graphs:

  • Excessively large or small fonts and symbols
  • Excessively thin or thick lines for curves and other graph features
  • Too many different variables being shown or varied at the same time
  • Having inappropriate axis selections (log versus linear, range, etc.)
  • Showing trends between data points with solid curves that indicate progressions or behavior that might not exist
  • Giving no indication of the magnitude of uncertainties or errors
  • Poorly captioned or labeled graphs (including units!)

In general, though, while most software packages are capable of creating good graphics, almost no program I've ever worked with defaults to a state that produces good graphics. They always require tweaking: either font sizes, or display ranges, or axes or symbol choices, and so on. Currently, I prefer to use matplotlib; others in my group have migrated to SciDavis.

  • $\begingroup$ Not sure what you mean by "Giving no indication of errors". can you clarify? Otherwise, good list. $\endgroup$ Dec 13 '11 at 19:53
  • $\begingroup$ Edited for clarity: "Giving no indication of the magnitude of uncertainties or errors." However, when the errors are too small to be shown, this can be indicated in the caption. $\endgroup$
    – aeismail
    Dec 13 '11 at 23:42

I've had reasonable success using the Mathematica package LevelScheme. It's execution model differs slightly from traditional Mathematica programming, so there is learning curve associated with its use. But, it is capable of providing fine control to plot generation which is difficult in plain Mathematica. Also, as a side package, there is a package for generating custom tick marks.

(Once the version with Mathematica v.8 support comes out it is going to be renamed as SciDraw.)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.