Sam Kaufman

Early attempts at reproducing the stepwise constant experiments from the paper with Gaussian data show some improvement for RF vs XGB, and, oddly enough, "extremely random" forests (ranger split criterion...

Here are the experimental results. I reproduce the stepwise-constant function in the shapr paper and run a 7-dimensional version of the 10d problem. Monte-carlo integration to get baseline shap. 400...

Thanks for the response, and for sharing the scripts. I will take a look - and perhaps continue the investigation. The TreeSHAP implementation I am using is the R 'TreeSHAP'...

Hi @salimamoukou! Very interesting! This clarifies some things re TreeSHAP - I initially had to struggle to understand why you claimed that TreeSHAP did not estimate your "reduced predictor" above...

Here is an example of the "elongated leaves" problem in 2 dimensions. The tree has 4 leaves (heavy lines) and the true function is constant on each leaf. The selected...

I think the SHAFF authors are thinking of using the partition of space learned by a forest model that's already been fit, simply for efficiency, but I think you're right,...

Thanks Paul, and thanks for your work on this very cool package. If I have time I will take a look at the code and see if a simple approach...

*edited* Just browsing - so half a thought. I wonder if the SHAPR "Empirical Conditional Distribution Approach/Conditional Inference Tree Approach" could use an adapted version of the kernel weighting already...

Adding similar requests #1582 and #1337

Thanks, that may be a neater approach. Do you have any idea how to make this work on spark? I see there are similar array functions, and will experiment with...