predict.warped_model: Predict from a warped fitted machine-learning model

In alexanderbrenning/wiml: Interpreting machine-learning models in transformed space

Predict from a warped fitted machine-learning model

Description

This function implements the prediction \textbf{w}\mapsto\hat{y}:=(\hat{f}\circ T^{-1})(\textbf{w}) by (back)transforming a feature vector w from the transformed space (e.g., principal components space) to the original feature space and then applying the fitted machine learning model's predict method to make predictions.

Usage

## S3 method for class 'warped_model'
predict(object, newdata = NULL, force_unwarp = TRUE, ...)

Arguments

object	A warped_model object, i.e. a warped fitted machine-learning model as created by warp_fitted_model.
newdata	A data frame with features in transformed space. This can be a warped_df object or just a plain data.frame. See force_unwarp.
force_unwarp	If TRUE, newdata will always be (back)transformed into feature space using the warper's inverse transformation. If FALSE, only objects of class warped_df will be backtransformed.
...	Additional arguments for the predict method of object$fit.

Details

newdata will normally be a data frame, not NULL; not all fitted models accept NULL. The most common behaviour is to return model predictions for the training data.

Value

The function's result depends on the behaviour of the invoked predict method.

See Also

warp_fitted_model() for creating a warped machine-learning model, and the predict method of your machine-learning model for details on additional arguments and the returned values.

Examples

### Create principal components warper for Maipo data set
### from Peña et al. (2015):
xvars <- c(paste("ndvi0", 1:8, sep = ""),
           paste("ndwi0", 1:8, sep = ""))
fo <- as.formula(paste("class ~", paste(xvars, collapse=" +" )))

sel <- seq(1, nrow(maipofields), by = 2)
dtrain <- maipofields[sel, ]
dtest <- maipofields[-sel, ]

wrp <- pca_warper(dtrain, xvars = xvars, yvar = "class")

fit <- randomForest::randomForest(formula = fo, data = dtrain)
fit
warped_fit <- warp_fitted_model(fit, warper = wrp)
class(warped_fit) # warped_model

# Create data frame with transformed test data:
wdtest <- warp(dtest, warper = wrp)

# Prediction using the transformed data and the warped model
# using the predict.warped_model method:
wpred <- predict(warped_fit, newdata = wdtest)
# ...should give the same results as prediction using the
# untransformed features and the original model using
# randomForest's predict method:
pred <- predict(fit, newdata = dtest)
all.equal(wpred, pred)

# Misclassification error rate on the test set:
mean(wdtest$class != wpred)

# Create profile plot along the first PC while
# keeping the other PCs = 0:
ngrid <- 100
newd <- matrix(0, ncol = length(wrp$Wvars), nrow = ngrid)
colnames(newd) <- wrp$Wvars
newd <- as.data.frame(newd)
newd$PC1 <- seq(min(wdtest$PC1), max(wdtest$PC1),
                length.out = ngrid)
# Using the predict.warped_model method:
prd <- predict(warped_fit, newdata = newd, type = "prob")

# Plot the predicted probabilities for varying PC1:
profile_plots <- function(pc = 1) {
  pc_name <- paste0("PC", pc)
  for (i in 1:4) {
    crop <- paste0("crop", i)
    if (i == 1) {
      plot(newd[, pc_name], prd[, crop],
           type = "l", lwd = 2,
           xlab = paste0("PC #", pc),
           ylab = "Predicted probability",
           main = "RF profile plots for crops 1-4",
           ylim = c(0, 1))
    } else {
      lines(newd[, pc_name], prd[, crop],
            lwd = 2, col = c("blue", "red", "magenta")[i - 1])
    }
  }
}

profile_plots(1)


### Repeat this for a linear discriminant analysis:

library("MASS")
fit <- MASS::lda(formula = fo, data = dtrain)
warped_fit <- warp_fitted_model(fit, warper = wrp)

# Prediction using the transformed data and the warped model:
wdtest <- warp(dtest, warper = wrp)
wpred <- predict(warped_fit, newdata = wdtest)$class
mean(wdtest$class != wpred) # better than RF!

# Create profile plot along the first PC:
prd <- predict(warped_fit, newdata = newd)$posterior
profile_plots(1)

alexanderbrenning/wiml documentation built on Sept. 29, 2023, 4:45 a.m.