22

I'm trying to explore the use of a GBM with h2o for a classification issue to replace a logistic regression (GLM). The non-linearity and interactions in my data make me think a GBM is more suitable.

I've ran a baseline GBM (see below) and compared the AUC against the AUC of the logistic regression. THe GBM performs much better.

In a classic linear logistic regression, one would be able to see the direction and effect of each of the predictors (x) on the outcome variable (y).

Now, I would like to evaluate the variable importance of the estimate GBM in the same way.

How does one obtain the variable importance for each of the (two) classes?

I know that the variable importance is not the same as the estimated coefficient in a logistic regression, but it would help me to understand which predictor impacts what class.

Others have asked similar questions, but the answers provided won't work for the H2O object.

Any help is much appreciated.

example.gbm <- h2o.gbm( x = c("list of predictors"), y = "binary response variable", training_frame = data, max_runtime_secs = 1800, nfolds=5, stopping_metric = "AUC")
Improve this question
5
  • have you tried: h2o.varimp(model)? Commented Dec 2, 2017 at 18:28
  • Yes, but that command gives the variable importance for both classes. Commented Dec 2, 2017 at 18:29
  • what is that you are referring to as 'linear logistic regression' . Would it be possible for you to elaborate further on what do you mean by 'variable importance for both classes' by an example and why it would not be same set of 'variable importance' for predicting both classes. Commented Dec 13, 2017 at 11:40
  • @GangeshDubey with 'linear logistic regression' in this case I simply refer to a regression for a binary variable. With respect to the 'variable importance for both classes' see the link above. Commented Dec 13, 2017 at 11:43
  • thanks, looked at the documentation for both h2o.gbm and h2o.grid , there appears to be no direct method to achieve it.in fact, had a look at the source code, you can validate that h2o.varimp returns a single value. Commented Dec 13, 2017 at 14:16

2 Answers 2

Reset to default
8

AFAIS, the more powerful a machine learning method, the more complex to explain what's going on beneath it.

The advantages of GBM method (as you mentioned already) also bring in difficulties to understand the model. This is especailly true for numeric varialbes when a GBM model may utilise value ranges differently that some may have positive impacts whereas others have negative effects.

For GLM, when there is no interaction specified, a numeric variable would be monotonic, hence you can have positive or negative impact examed.

Now that a total view is difficult, is there any method we can analyse the model? There are 2 methods we can start with:

Partial Dependence Plot

h2o provides h2o.partialplot that gives the partial (i.e. marginal) effect for each variable, which can be seen as the effect:

library(h2o) h2o.init() prostate.path <- system.file("extdata", "prostate.csv", package="h2o") prostate.hex <- h2o.uploadFile(path = prostate.path, destination_frame = "prostate.hex") prostate.hex[, "CAPSULE"] <- as.factor(prostate.hex[, "CAPSULE"] ) prostate.hex[, "RACE"] <- as.factor(prostate.hex[,"RACE"] ) prostate.gbm <- h2o.gbm(x = c("AGE","RACE"), y = "CAPSULE", training_frame = prostate.hex, ntrees = 10, max_depth = 5, learn_rate = 0.1) h2o.partialPlot(object = prostate.gbm, data = prostate.hex, cols = "AGE")

enter image description here

Individual Analyser

LIME package [https://github.com/thomasp85/lime] provides capability to check variables contribution for each of observations. Luckily, this r package supports h2o already.

enter image description here

Improve this answer
Sign up to request clarification or add additional context in comments.

Comments

1

You can try h2o.varimp(object)

Improve this answer

Comments

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.