Axes first, story second

Your eye goes straight to the lines and bars. Train it to go to the axes first. The same data can tell very different stories depending on the axes.

Four questions for every figure

1. What is on each axis, and in what units?
2. Where does the y-axis start? A bar chart that starts at 90% instead of 0% makes a 1-point gap look huge.
3. Is the scale linear or log? On a log scale, equal steps mean equal multiples (10, 100, 1000), not equal amounts.
4. What is one point? One run, an average of five, one person, one dataset?

The cut-off axis

ExampleTwo bars: 92% and 93% accuracy. With the axis from 0 to 100, they look almost the same, which is honest. With the axis from 91 to 94, one bar looks three times taller. Nothing in the data changed.

Cutting the axis is sometimes reasonable, to show small but real differences. But you should notice it, and judge the size of the gap by the numbers, not the picture.

Log scales

Log scales are common in ML for training curves, dataset sizes and compute. They are useful, but they hide big differences at the top. A line that looks "nearly flat" on a log axis can still be gaining a lot in real terms.

Read the caption

A good caption says what each point is, how many runs, and what the error bars mean. If the caption does not say, that is a gap in the evidence, and worth a note.

RememberRead axes, scale and caption before the data. Judge differences by the numbers, not by how big they look.