[Meta-issue] Improve plotting API for consistency about shap HOT 8 OPEN

From the prototype implementation of this new approach on #3523, I noticed a possible pitfall to avoid. Often the plotting functions are called like this:

def foo():
    f = plt.figure()
    shap.plots.bar(..., show=False)
    return f

To avoid breaking this pattern, we should avoid calling plt.figure() during the plotting functions, even if ax=None. Instead, we should continue using plt.gcf(), which will effectively be the same as plt.figure() if no current figure exists. This will keep the new code consistent with the library's previous behaviour.

from shap.

Hi, I'm new to contributing and I would love to help out with a few of these. I'm happy to pick up the beeswarm plot.

from shap.

Hi, I'm new to contributing and I would love to help out with a few of these. I'm happy to pick up the beeswarm plot.

@chriscave fantastic, great you to have you as a contributor. Give us a shout if you'd like any pointers - for example, feel free to tag me in a draft PR.

from shap.

For Jupyter notebooks, this approach works great! I'm using SHAP in a web application. Here's some sample code I wrote:

feature_importance = np.mean(np.abs(shap_values.values), axis=0)
feature_names = shap_values.feature_names
feature_names, feature_importance = zip(*sorted(zip(feature_names, feature_importance), key=lambda x: x[1], reverse=True))

I then pass this information into my display layer to render as a real-time HTML chart (in my case, chart.js). Has there been any consideration of an approach where:

• Some of the logic (e.g., feature importance) is refactored into Explanation objects (or maybe somewhere else; I'm not that familiar with the code)
• Making the various plot functions just do display instead of some of the analysis computation?

This would make it much easier for applications that embed SHAP, and also simplify the logic in the plot functions. But maybe I'm missing something. Happy to chat live if it would be helpful, and also willing to help contribute in this area.

from shap.

Hi, I have a question regarding more complex plots like the heatmap. I looked at the code and it felt like a lot of hacky things are going on to fit the heatmap, the horizontal barplot, the f(x) line plot and the colorbar into one "Axes".

What is the opinion on treating the heatmap as a figure level plot with axes for the sub-elements? The improved API style could be provided by giving the option to pass a figure object, and by adding explicit labels to the axes, it could be easier to modify downstream.

from shap.

What is the opinion on treating the heatmap as a figure level plot with axes for the sub-elements? The improved API style could be provided by giving the option to pass a figure object, and by adding explicit labels to the axes, it could be easier to modify downstream.

The current approach has the advantage that all plots are axes-level plots, so they can have a consistent API. I would prefer to keep the API consistent and keep the visual appearance of the plots the same whilst we modify them. However, I'd be open to that change in future if it makes the implementation significantly clearer.

It would be unusual to accept a fig as an input argument, as defining which plots should go on which inner axes may become rather complex. It is probably simpler to start by accepting the ax directly.

from shap.

The current approach has the advantage that all plots are axes-level plots, so they can have a consistent API. I would prefer to keep the API consistent and keep the visual appearance of the plots the same whilst we modify them. However, I'd be open to that change in future if it makes the implementation significantly clearer.

It would be unusual to accept a fig as an input argument, as defining which plots should go on which inner axes may become rather complex. It is probably simpler to start by accepting the ax directly.

Thank you for the reply and input. I can see the value of having a consistent API.

I am working on a multi heatmap plot for a multiclass classification problem and would like to try a few things, making it possible to modify the plot a bit (e.g. removing the colorbar, adding a sample annotation) and I'm not entirely sure if that works with the ax level API. @connortann Do you mind, if I tag you in a future PR?

from shap.

Sure, we're always happy to have new contributors!

from shap.