Shap: nsamples for TreeExplainer

[cbeauhilton]

I'm using a TreeExplainer to calculate SHAP values for a large dataset (~1.7m rows, ~170 cols), model generated with LightGBM. Is there a preferred method to take nsamples for the background to speed up the process? Or is this not possible with the TreeExplainer?

[slundberg]

The nice thing about TreeExplainer is that it does not need to take any samples (unless you use the feature_dependence="independent" option). If you are trying to get a summary of your data I would just explain a random subsample of 10k and then plot whatever you want. If you really need to explain every sample very quickly you could try using approximate=True which will probably be a reasonable approximation if you have lots and lots of trees.

[cbeauhilton]

Awesome, thanks, as always.

I was getting similar SHAP results when running debugging versions of the experiments with 10-20k samples as with the full 1.7m, I just wanted to make sure this was theoretically sound (enough).

So I will move forward with small samples for explanation, then probably just let the thing run overnight for the publication, more as an exercise in completeness than anything else.

[slundberg]

Sounds good! Just beware of putting 1.7m points into the plotting functions. They might not like that many points...