Add a text for displaying values on histogram.

[lionelkusch]

For the histogram, I add text for displaying the percentage of each bar, the same as for values_count. I try to reuse the idea of display value for displaying on the histogram.

This pull request was made for fixing the issue #1454.

[Vincent-Maladiere]

Hey @lionelkusch thank you for this PR! #1454 is kinda tricky and we're not sure how to move forward.

Your suggestion is interesting, but my understanding is that we'd rather have the height of each vertical bar rather than the percentile of their distribution. Additionally, we'd also like to have the total count for the categorical distribution (displayed horizontally). We need to see whether adding these figures would take too much space to be useful or not.

We can use this PR to explore different options.

[lionelkusch]

This PR is quite limited to add information to the histogram is the same spirit as for the value_counted.
I pushed further in PR #1506 . For adding the height of the bar.
Tell me what final form you want.

[lionelkusch]

Addressing your remarks:

The problem when you combine the two in the same graphic is that the text will be longer than the columns. This won't look nice and it will be difficult to automatically add the text to the graphic.

If the text crossing the bar edge doesn't look good, we can choose to always have the text outside of the bar, increasing the figure size slightly if necessary.

Adding the number of elements creates some variability and creates a difficult problem for the position of the text relative to the bar.

I think this can be addressed by controlling the number of digits in the format using custom rules, like full representation when the entries are smaller than 10,000 and scientific notation otherwise. We would have an upper bound on the size of the text.

---

What do the others think?

I agree on all points here and I would also prefer having both numbers visible at the same time.

Adding the number of elements creates some variability and creates a difficult problem for the position of the text relative to the bar.

How about adding the % to the right (the top) of the plot, aligned with each bar? With a fixed format (like 3 significant digits) the space should remain mostly constant, while the actual value is still represented in the same way as it is now. If the % sign takes too much space, it could be added above the column of numbers as an annotation.

Originally posted by @rcap107 in #1506 (comment)

I put back the comments of the PR #1506 about the graphic representation.

[rcap107]

I quite like the current version, thanks a lot @lionelkusch!

I just realized is that we might want to remove the % numbers for numerical features, because bins for numerical values are defined somewhat arbitrarily, so the % numbers for them are not really meaningful.

[rcap107]

Hey @lionelkusch, do you think you finish this PR? As we discussed, the current version is good, the only change would be removing the % from the numerical columns

I hope there won't be too many conflicts to deal with

[lionelkusch]

@rcap107 I am sorry, I forgot to do the modification.
I did it. I think that the PR is read.

[rcap107]

thanks a lot! I'm wondering if the plots would look better without the right and top spines, like the current version 🤔

[lionelkusch]

I propose you this solution:

[GaelVaroquaux]

Is there a place where I can look at a full HTML version, to interact a bit with it?

[rcap107]

Is there a place where I can look at a full HTML version, to interact a bit with it?

I was comparing the table report in the homepage of the generated docs with that in the live version

[rcap107]

I propose you this solution:

I think it's ok, I was worried that the tick might be read by someone as a "minus" or "negative", but it doesn't really make sense for percentages

[lionelkusch]

I removed the tick line; I think it's not needed it for the readability of the figure.

[GaelVaroquaux]

It's quite OK, actually:

My concern is: information overload. The more we put, the harder it is for a user to focus and to get the message. It brings us to the gold rule of communication: less in more.

But you've managed to keep it really light.

An ideal option, would have been to have a "hover": information that appears when the mouse goes over the relevant graph. But it's hard to do with our current toolkit.

I'm +1 for merge on this. Congratulations!!

[rcap107]

Hey @lionelkusch, I keep on losing track of this issue 🙈 my bad

We can merge after removing the barplot labels for the numerical features, like here

The bins are decided arbitrarily, so the size of each bin is not particularly informative.

(I thought I had already mentioned this here, but I guess I forgot)

[lionelkusch]

Are you sure that you want to remove the label?
I ask the question because it seems to me that this was the starting point of the PR, as described in the issue #1454.

[rcap107]

Are you sure that you want to remove the label? I ask the question because it seems to me that this was the starting point of the PR, as described in the issue #1454.

#1454 mentioned both categorical and numerical features, but given that the distribution of numerical features is generated with "best effort" bins, the actual value of the bins isn't very meaningful, so I think it can be removed.

For categorical features, each category is well delimited so it makes sense to have both the number of occurrences and the %, so the current version is good and should remain as you wrote it.

[lionelkusch]

done

[rcap107]

Looks good to me, thanks a lot @lionelkusch. It was a bit convoluted (my fault for that), but the result is a nice addition to the TableReport.

[MA-inria]

Thank you all for merging this PR