Seaborn: Weighting the samples of a kdeplot

[arvoelke]

Related:

• #783
• https://stackoverflow.com/questions/31703149/weights-option-for-seaborn-distplot
• https://stackoverflow.com/questions/30344568/how-to-pass-weights-argument-to-seaborns-jointplot-or-the-underlying-kdeplot

It would be nice to be able to create a kdeplot where a weight can be assigned to each sample. Even though there is a hist_kws argument, it is not taken into account with the kernel density estimate.

Here's a hypothetical use-case: you have data recording the (income, donation) of a number of people to some charity, and now you want to visualize the univariate distribution of donation (y-axis) against income (x-axis). If you try to do this with a kdeplot, currently, the amount donated cannot be included in the density, i.e., each donation is treated uniformly.

My work-around has been to use the weights+observations to generate a million-or-so samples according to that distribution, and then pass those samples to the kdeplot. This tends to be slow/noisy, and awkward to implement.

By the way... thank you to the community for the very useful toolkit. I have used it in a recent publication (http://ieeexplore.ieee.org/document/7378880/) and cited seaborn accordingly. Incidentally, I had to use the above work-around of generating proxy samples for that paper in order to have the kdeplot weighted correctly.

[mwaskom]

The KDE is fit with scipy, which doesn't seem to support weights.

Your hypothetical use-case seems like something you want to be doing with a lowess regression, not a weighted KDE, anyway.

[cstich]

It looks like scipy will with the 1.2.0 release support weights for KDE. The relevant pull request is here. Any chance of opening this issue again?

[MaxGhenis]

Agree I think this should be reopened, given it's being worked on in #1747.

[mwaskom]

Closed in #2104 with the addition of weights to kdeplot as part of a very thorough rewrite.