SHAPR for custom defined models

[mehwish2021]

Hello All,
I am using DeepCC model in which I want to apply SHAPR to get the shaply values .
I defined the functions as described in here https://cran.r-project.org/web/packages/shapr/vignettes/understanding_shapr.html , but its still giving me the error that SHAPR is not defined for this model.

I have defined these functions get_model_specs() and predict_model() as mentioned in the vignette but I am confused how and where are those functions called?
Please help as it has taken much time and I am unable to resolve this issue

[martinju]

Hi! Please use the github version of the shapr package.

[mehwish2021]

Thanks for your prompt response Martinju.
A little more clarification, do you mean I have to add these functions in the code of shapR package?? Can you please guide in which file I need to add these functions ?
I would really appreciate you response

[martinju]

No, the package on CRAN is outdated (and will be replaced by the main at GitHub) soon. It does things a bit differently. I suggest you install the github version of shapr instead, modify your code accordingly, try to use that to explain your custom model instead. The procedure is explained her:
https://norskregnesentral.github.io/shapr/articles/understanding_shapr.html#explain-custom-models
(which is slightly different from the procedure with the version on CRAN)

If you still got issues after that, let me know.

[vmombo]

Hello All, I have a Keras model, is it possible to make it work with it?

[martinju]

Hello All, I have a Keras model, is it possible to make it work with it?

Yes. Please see the main vignetten on the pkgdown site for instructions.

[vmombo]

Thanks. I tried and that works well on model I built with the Boston Dataset. However with one of my real model I get this

`explanation <- explain(
model = finalModel,
x_explain = sdt.test,
x_train = dt.test[1:100, ],
approach = "empirical", # Choose explanation approach
phi0 = phi0, # Specify baseline value
predict_model = predict_model.EnsembleModel,
get_model_specs = get_model_specs,
#max_n_coalitions = 2^10, # Reasonable limit
verbose = "progress"
)

Success with message:
max_n_coalitions is NULL or larger than or 2^n_features = 3.86856262276681e+25,
and is therefore set to 2^n_features = 3.86856262276681e+25.

Error in .makeMessage(..., domain = domain) :
argument is missing, with no default`

I am trying to understand this error by looking on the source code I still don't understand. Any suggest about where I should investigate.?
PS: I use the version from the main of this github repo.

[martinju]

It seems you have an enormous number of features? (83 or 84 features?)

I have not seen this error message before, but in any case, you should first try to reduce the number of features to see if it works then.
Also, set max_n_coalitions to 500 or something to reduce runtime when testing.

[vmombo]

Yes, excellent. But I've dug into the code a little bit to have like a traceback of my error.

I think apart from the big number of features. my error is coming from the warning in the code. Especially for me . In the explain function, we call setup and in the function check_computability my input leads me to this condition:

if (isFALSE(is_groupwise) && n_features > 30) { warning( "Due to computation time, we strongly recommend enabling iterative estimation with iterative = TRUE", " when n_features > 30.\n", ) }

However a warning written this way produce an error because of the "," at the end. You can reproduce my error just by calling this line :
warning( "Due to computation time, we strongly recommend enabling iterative estimation with iterative = TRUE", " when n_features > 30.\n", )

[martinju]

Oh, thanks a lot for catching that bug! I will fix it ASAP

[martinju]

@vmombo I just merged the fix to main

[vmombo]

thank you very much. I have just seen your comment

[vmombo]

Hi @martinju ,

I've 85 variables to explain for some predictions, I've tried several times but I got some crashes with the following error : R(27812,0x2033e3840) malloc: *** error for object 0x600001bdc080: pointer being freed was not allocated R(27812,0x2033e3840) malloc: *** set a breakpoint in malloc_error_break to debug,

I’m not sure if this is due to my computer's capacity, but I suspect it might be (so many coalitions). I found a "walk around" which is by grouping variables as you did in the paper.

However and most importantly, I would like to know the following:

• In the grouped approach, what exactly is being explained in the predictions?
• Is it the mean value of the group?

Thanks in advance, for your help