Don't Connect the Data Points on your Scatterplots

Posted on Dec 29, 2023

Now today, please permit me to rant for a second about connecting the dots when you plot data. Stop it! I mean, probably. Ask yourself: “does it make sense to interpolate these data points?” The answer is probably “no” and you should just do a scatterplot with individual points.

There are counterexamples: maybe you plot an ROC curve for a binary classifier. It is literally possible take linear combinations of classifiers to get the interpolated properties. Go right ahead.

But most other measurements, not so much. Your algorithm does not have a runtime on graphs with 42.5 vertices. I guess if you’re plotting averages and the instances come from a distribution of sizes: maybe, but you see what I mean. And if you can, maybe don’t take averages: give me a box plot or even scatter all your measurements.

Stay in school, don’t do chart crime.