Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

partial() function giving inconsistent results when feeding multiple pred.var variables #130

Open
rahualram opened this issue Feb 20, 2023 · 1 comment

Comments

@rahualram
Copy link

rahualram commented Feb 20, 2023

Hello,

I've been running into an issue with the partial function when using it across multiple features with a custom grid table. I've noticed that the partial function works well with a single row in the train dataset but doesn't seem to give sensible results when you add more rows and vary multiple features. Here is an example:

# Load required packages
library(pdp)
library(xgboost)

# Load the data
data(pima)
X <- subset(pima, select = -diabetes)
y <- ifelse(pima$diabetes == "pos", 1, 0)

# Parameters for XGBoost model
param.list <- list(max_depth = 5, eta = 0.01, objective = "binary:logistic", 
                   eval_metric = "auc")

# Fit an XGBoost model
set.seed(101)
pima.xgb <- xgb.train(params = param.list, 
                      data = xgb.DMatrix(data.matrix(X), label = y), 
                      nrounds = 500)

grid <- data.table(mass = c(20,40), age = c(30, 50))

all_predict <- cbind(X, prediction = predict(pima.xgb, xgb.DMatrix(data.matrix(X), label = y)))

person_1 <- partial(pima.xgb, pred.var = c("mass", "age"), train = X[1,], pred.grid = grid)

person_2 <- partial(pima.xgb, pred.var = c("mass", "age"), train = X[2,], pred.grid = grid)

both_ppl <- partial(pima.xgb, pred.var = c("mass", "age"), train = X[c(1,2),], pred.grid = grid)

results <- merge(person_1, person_2, c("mass", "age"))
results <- as.data.table(merge(results, both_ppl, c("mass", "age")))
results <- results %>% rename("person1" = "yhat.x", "person2" = "yhat.y", "both" = "yhat")
results[,avg := (person1+person2)/2]

view(results)

The above shows that when using the partial function for more than one row, the result deviates from the average prediction of each row. I've had a look into what is causing this and at first I thought it was due to the differences in xgb.DMatrix verse data.matrix but I ended up replicating the error by pulling apart the pardep function a bit. I noticed that the following line of code in the pardep function seems to be causing the issue when pred.var is more than one feature:

temp[, pred.var] <- pred.grid[i, pred.var]

It seems to be that the assignment of the pred.grid to a data.matrix which is then subset to a single row somehow transforms this into a single array which is used to populate temp by column rather than row. Using the example above, my dataset is meant to look like

          Mass        Age
Row 1:      20            30
Row 2:      20            30

when running the first grid point (i.e i =1 in the foreach loop).
But what I'm seeing when stepping through is:

           Mass        Age
Row 1:      20            20
Row 2:      30            30

Is this a known issue or am I doing something wrong when feeding in the parameters to partial?

@bgreenwell
Copy link
Owner

Hi @rahualram, thanks for the note. I'll look into this shortly!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants