Calculate shap values only for the selected features

[mnwright]

While the components were correct, the SHAP values were all 0 for the non-selected features, as in this example from #13:

library(glex)
library(xgboost)

x <- as.matrix(mtcars[, -1])
y <- mtcars$mpg

xg <- xgboost(data = x, label = y, params = list(max_depth = 4, eta = .1), nrounds = 100, verbose = 0)

glex(xg, x, features = c("cyl", "disp"))

But I wonder, should we even calculate the SHAP values? They are not correct SHAP values anyway because we just ignored some of the features. That means, in the above example and

res1 <- glex(xg, x, features = c("cyl", "disp"))
res2 <- glex(xg, x)

the components res1$m and res2$m are the same, but the SHAP values res1$shap and res2$shap not.

[jemus42]

We started carrying around a $remainder term for more or less that reason, no?
Feels like it would need a huge disclaimer in the docs though.

[jemus42]

After discussing this a little offline I think the safest choice is to omit the $shap part in cases where we artificially restrict the number of features in glex(), and we should maybe consider whether we might want to do the same thing when max_interactions is restricted as well.

[jemus42]

Picked up and extended in #39