Improving Diagrams

Today I'll give a little example on how to improve a diagram for a talk. For printing there might be other considerations (such as choosing colors that print well in black and white) but today we'll concentrate on diagrams that you want to show during a presentation.

Let's take some of my data where we made WiFi measurements using a tablet and a smartphone and measured reception at different locations. If you simply generate a standard diagram with Excel from that data you might end up with a slide looking like this:

A terrible slide

Of course, this is utterly terrible!

Let's take a closer look at the why. First, those were measurement points, so using a line diagram is wrong anyway; plus, all those lines are totally confusing. Second, what we really would like to show is a comparison between tablet and phone reception and having the data in two different diagrams below each other makes this task virtually impossible.
The only thing we can really see from this diagram is that there is no especially strong or weak channel (we measured channels 1, 3, 5, 7, 9, and 11 for each of the measurement points with each device).

As a first improvement we will show the data we want to compare in one combined diagram and eliminate the lines. We end up with something like this:

With all those symbols it's still very confusing and we can't really compare tablet and phone reception. When we look at the numerical data we can read that the tablet has somewhat better reception, so let's make this clear graphically. An easy way to do this is to use one color for each group of data. Here, we choose orange for all tablet measurements and blue for all phone measurements. Orange and blue are complementary colors which look nice together and there is also no problem with colorblindness (most colorblind people are red/green colorblind with blue/green colorblindness being the second most common). So, our next improvement looks like this:

This already gives a pretty good picture but still has some minor technical issues. First, symbols not only have a fill color but also a frame color and those lines from the frames irritate the eye. Second, some symbols (phone measurements for channel 3 and 5) are so small that they practically consist only of frame color and don't have any visible fill color. Thirdly, many symbols are on top of each other but it's hard to determine where.

So we will do the following: we will use symbols without frames and make symbols 50% transparent. Since we have observed that the channel does not play a role and there is no story there to tell we will also make all symbols for each group of data the same. (Basically, we will treat the different channels like we would treat different measurements taken at different times.) We choose circles for all tablet measurements and squares for all phone measurements. We can now eliminate all surplus legend entries, too. We move the remaining legend entries to the bottom of the diagram and use the increased horizontal space to make the diagram larger so there is more whitespace between columns. Our improved slide now looks like this:

Pretty good!

Now we can really see what's going on and all distractions have been eliminated. Note that when two measurements fall on top of each other, the resulting point looks like a darker version of the symbol as the semi-transparent symbols stack up on top of each other.

Since this is still technically a line diagram (just with disabled lines) we can also use custom annotations for our horizontal axis. This is not possible when using a point diagram in Excel (which is a bug IMO).

But what if you wanted to also show which channel was which in this diagram? You could use eg. shadings of the same color and might end up with something like this:

A try, but not really useful

I would not recommend this, though, because light colors plus transparency might render invisible when projected and people might intuitively associate color brightness with signal strength. Also, since channel influence is random, there is no story here to tell anyway. Therefore, I would not clutter up my diagram for this non-info. If there were an especially strong or weak channel (or small group of channels) I would consider using similar colors - eg. using orange and red (both "warm" colors) for the tablet measurements and two "cold" colors (maybe blue and grey) for the phone measurements to show their respective strong or weak channels.

One thing we might want to do is add a trend line. For this we calculate the average over all channels at each measurement point. Graphically, we will use a line instead of single points to separate the average better from the actual measurement points. Do we have to add a legend entry for the trend lines? IMO, we can go without because, a) we will show this slide during a talk, so we are there to explain it anyway and b) even if someone comes across the slide on its own it is still pretty obvious what the lines mean so I would go with intuitive clarity over extra legend clutter in this case.

What we notice here is that at point H the average of the tablet measurements is below the phone measurements. This is only possible if at this point a lot of tablet measurements are very low so the circles end up being hidden by the squares. Maybe it's time to rethink our choice of symbols to avoid them obscuring each other so much. So for a final version I chose triangles for the phone measurements so the circles can shine through a bit better:

Quite an improvement over the first version of our slide!

Find more on giving good talks HERE.

If you want to see how you can do something similar in gnuplot, see this post.