Small demo of shap-like feature

[GaelVaroquaux]

Fixes #89 to make shap-like features more easy to discover

[JingangQu]

Fine for me!

[marineLM]

For coherence, I think it would make sense to present it as a separate tutorial rather than include it in the Getting Started page.

[GaelVaroquaux]

I think it would make sense to present it as a separate tutorial rather than include it in the Getting Started page.

I can do it if you want.

The benefit would be to make it more visible.

The drawback is that long term a proliferation of small examples crowds a documentation and gives a daunting page of examples. Of course, the solution would be to restructure the documentation when this happen, but experience shows that this does not happen, because it's a lot of work and people are seldom excited about this.

@marineLM : tell me what you want me to do.

[dholzmueller]

What about we make a separate SHAP tutorial (I also thought of doing it this way) but then we also put the other extensions (unsupervised learning) in the same tutorial later on?

[GaelVaroquaux]

OK. One thing to keep in mind is that time is limited, and having something is better than having nothing.

I can dedicate another 10 or 20mn to this, not more. The current status is that our shap features are not exposed at all. People won't find it, and right now the conversation on interpretability and TFM is focusing on the packages that expose shap features.

[GaelVaroquaux]

OK, I'm making an editorial decision (because I can try and work on this now, and will not later):

The example that I did is a bit crappy in term of interpretability, because the data is synthetic, and the column names don't anything.

I'll try to make an example using the data in https://scikit-learn.org/stable/auto_examples/inspection/plot_linear_model_coefficient_interpretation.html
I'll see if I have time (no promises, I might get interrupted)

[GaelVaroquaux]

I have error, related to numba, that I don't fully understand.

Might be legit, @JingangQu

[JingangQu]

I have error, related to numba, that I don't fully understand.

Might be legit, @JingangQu

@GaelVaroquaux, I found out the problem. the following code in your example

X = pd.get_dummies(X, drop_first=True)

yields boolean columns with True / False. In the shap file, I've used X_np = np.asarray(X_test) producing an object dtype array. SHAP's masker passes this to a numba @njit function which can't handle object arrays.

I fix this problem via X_np = np.asarray(X_test, dtype=np.float64).

btw, your shap example takes too much time (about 15 minutes on my macbook with m3 pro). We need a lightweight example. I'll do it.

[GaelVaroquaux]

Thanks for the fix, @JingangQu . I think that this look great.

IMHO, we should merge. It's very useful, and I want to use it on social media :)

[JingangQu]

Thanks for the fix, @JingangQu . I think that this look great.

IMHO, we should merge. It's very useful, and I want to use it on social media :)

Just wait for a moment. There is an issue in the toturial figures. I am fixing it. After that, I will merge it