Many people will agree with the Less is more gif. However, I read someone states that in scientific chart, you should not remove axis or use direct label. But I don't know why, I see no reason to avoid this. This is my chart with the less-is-more philosophy.
It depends on your goals.
Who is reading the graph? In print or on screen? Web or PDF? Are the numbers important, or just the relative magnitudes?
I'd say that, in this kind of plot, the number labels are fine and the axis label is probably unnecessary. That's because there are only a few bars and the scale is linear.
But in general these questions need to be decided on a case by case basis, and numerical labels can get ugly quickly. Having too much information is almost as bad as having too little. For example, it's totally counterproductive to report regression coefficients to the 100s place unless your measurements are that precise and that precision actually matters. The same is true for plots, where the spatial variation along an axis is more than sufficient for making comparisons.
The reason I like the numbers here is that comparing distances along an axis takes eye movement and therefore effort: you need to constantly "flip" back and forth between the plot area and the axis. Dropping the axis and labeling the bars directly saves the reader that effort.
Dogmatic adherence to Tufte and Cleveland is a sign that you either didn't read them or didn't understand them.
Here's some sage advice from Jane Miller:
Data labels are typed numeric values adjacent to the pertinent slice, point, or bar in a chart ... To keep charts simple and readable, use data labels sparingly. Again, the main advantage of a chart is that it can illustrate general levels or patterns, which will be evident without data labels if your chart has adequate titles. Complement the general depiction in the chart with your text description, reporting exact values of selected numbers to document the patterns ... If your audience requires exact values for all numbers in the chart, replace the chart with a table or include an appendix table rather than putting data labels on each point. Reserve data labels for
- reference points; and
- reporting absolute level associated with a pie or stacked bar chart, such as total number of cases or total value of the contents of the pie
I don't like the direct labeling without the y-axis because many times the y axis gets truncated.
For example, compare two plots that show the same data. In the first one, if there was no y axis, at first glance, I would say that there was quite a bit of variation across the five groups. To make the correct inference, I would have to read and compare five labels, which is more cognitively demanding. At that point, you might as well use a table. When I see that the y-axis doesn't start at zero, it makes my Spider sense tingle. Removing the axis forecloses using that heuristic and creates more for my brain to do.
The second graph is more honest, but the labels would add clutter. One possible reason to use them is that someone might want to do some calculations based on those numbers. I view gridlines as a nice compromise between the labeling and less is more approach. In the digital realm, labels that appear when you click or hover over the bars are another option.
Another example of axes that you don't want removed is logarithmic one, where you want labels on the original scale.