Error during explain with H2O GBM/XGB models - "NA/NaN/Inf in 'x'"

[dkincaid]

I'm trying to use the package with an H2O xgboost model (I've also tried it with GBM and get the same thing. The error is:

Error in glm.fit(x = x_fit, y = y, weights = weights, family = gaussian()) : 
  NA/NaN/Inf in 'x'

Here is the code I'm running:

explainer <- lime::lime(as.data.frame(wellnessTrain), mdl)

explanation <- lime::explain(as.data.frame(wellnessTest),
                       explainer, n_labels = 1, n_features = 2)

This is caused by having some NA values in the data frame, but I thought that this had already been fixed in issue #8. I verified this by removing the three columns that have NA values as a test. These NA values are meaningful and H2O's GBM and XGBoost handle them by creating a category for the missing value after binning the unmissing feature values. Is there any easy fix here?

[dkincaid]

I have found that I can use feature_select="tree", but none of the other possible types.

[thomasp85]

Where do you get wellness data from? and how is your model made? I would need it to reproduce your example...

[dkincaid]

Unfortunately I can't share that data, but let me try to put together a reproducible example with some publicly available data. I was hoping that something I was doing would look wrong.

[thomasp85]

Not obviously so - I’ll look into it if you can make a reprex

[dkincaid]

Here is a reproducible example using the Iris data. I'm showing a successful run with the full iris data frame and then run the same thing against a data frame where I randomly set some values to NA. Hopefully this gives you something to work with. I appreciate you taking a look at it.

# Create a data frame from the Iris data and randomly set some values to NA
myIris <- purrr::map_df(iris[,-5], function(x) {x[sample(c(TRUE, NA), prob = c(0.8, 0.2), size = length(x), replace = TRUE)]})

myIris <- cbind(myIris, Species=iris$Species)

library(h2o)
h2o.init()

# First show that it's successful without any missing data
full_iris_frame <- as.h2o(iris)
full_mdl <- h2o.gbm(training_frame = full_iris_frame, y = "Species")

full_explainer <- lime::lime(dplyr::select(as.data.frame(full_iris_frame), -Species), full_mdl)

full_explanation <- lime::explain(dplyr::select(as.data.frame(full_iris_frame)[1:4,], -Species),
                             full_explainer, n_labels = 3 , n_features = 3)

# Now try to run it on the data that has some missing values
iris_frame <- as.h2o(myIris)
mdl <- h2o.gbm(training_frame = iris_frame,
               y = "Species")

explainer <- lime::lime(dplyr::select(as.data.frame(iris_frame), -Species), mdl)

explanation <- lime::explain(dplyr::select(as.data.frame(iris_frame)[1:4,], -Species),
                             explainer, n_labels = 3 , n_features = 3)

[thomasp85]

So, the support for NA implemented earlier were only considering NA values in the training data - not NA values in new observations to explain. I've just pushed an update that ignores NA columns in new observations so that you don't get the error.

[dkincaid]

That is fantastic! Thanks for such a quick fix. I really love this package.