predict from ranger (more modern RF)

[m-r-munroe]

I'm trying to approximately reproduce the figure from the figure1 from the arxiv paper. The CCF seems to work well with 35 trees.

This is what the reference figure looks like:

A decent RF tool that isn't 20 years old is ranger.

To get a decent plot using the canonical correlation forest I use this:

require(pacman)

p_load_gh("jandob/ccf")
p_load(tidyverse, 
       ranger,
       ggplot2,
       GGally)

#parameter
num_trees <- 35

# load sample dataset
data(spirals)

#split data
d_train <- spirals[1:1000, ]
d_test <- tail(spirals, 1000)



## make ccf on the data
ccf_f <- canonical_correlation_forest(class ~ ., d_train, ntree = num_trees)

# plot the decision surface of the classifier
ccf_plot <- plot_decision_surface(model = ccf_f, 
                                  X = d_test[, c("x", "y")], 
                                  Y = d_test$class,
                                  title = str_c("CCF with ", num_trees, " trees"))

print(ccf_plot)

to make this:

When I use the following code, it doesn't work so well:

rngr_fit <- ranger(class ~ ., data = d_train, num.tree = num_trees)
rngr_plot <- plot_decision_surface(model = rngr_fit, 
                                  X = d_test[, c("x", "y")], 
                                  Y = d_test$class,
                                  title = str_c("Ranger with ", num_trees, " trees"))

print(rngr_plot)

The output is this:

> rngr_fit <- ranger(class ~ ., data = d_train, num.tree = num_trees, classification = TRUE)
> rngr_plot <- plot_decision_surface(model = rngr_fit, 
+                                   X = d_test[, c("x", "y")], 
+                                   Y = d_test$class,
+                                   title = str_c("Ranger with ", num_trees, " trees"))
Error in as.data.frame.default(x[[i]], optional = TRUE, stringsAsFactors = stringsAsFactors) : 
  cannot coerce class ‘"ranger.prediction"’ to a data.frame
> print(rngr_plot)
Error in print(rngr_plot) : object 'rngr_plot' not found

It does not like the output of the ranger prediction. This is the structure of the prediction object.

> str(predict(rngr_fit, data = d_test))
List of 5
 $ predictions              : num [1:1000] 3 3 2 3 2 3 1 2 2 1 ...
 $ num.trees                : num 35
 $ num.independent.variables: num 2
 $ num.samples              : int 1000
 $ treetype                 : chr "Classification"
 - attr(*, "class")= chr "ranger.prediction"

This is the structure of the ranger$predictions object:

> str(predict(rngr_fit, data = d_test)$predictions %>% as.character())
 chr [1:1000] "3" "3" "2" "3" "2" "3" "1" "2" "2" "1" "1" "3" "2" "3" "3" "3" "1" "2" "3" "1" "1" "2" "2" "1" "3" "1" "3" ...

it looks a lot like the output of the ccf output object.

> str(predict(ccf_f, newdata = d_test))
 chr [1:1000] "3" "3" "2" "3" "2" "3" "1" "2" "2" "1" "1" "3" "2" "3" "3" "3" "1" "2" "3" "1" "1" "2" "2" "1" "3" "1" "3" ...

When I look at what R is doing wit the plot_decision_surface I get this:

function (model, X, Y, title = NULL, interpolate = FALSE, ...) 
{
   data <- data.frame(x = X[, 1], y = X[, 2], z = Y)
   x_min <- min(data$x) * 1.2
   x_max <- max(data$x) * 1.2
   y_min <- min(data$y) * 1.2
   y_max <- max(data$y) * 1.2
   resolution <- 400
   grid <- expand.grid(x = seq(x_min, x_max, length.out = resolution), 
      y = seq(y_min, y_max, length.out = resolution))
   predictions <- predict(model, grid, ...)
   data_raster <- data.frame(x = grid$x, y = grid$y, z = predictions)
   plot_object <- generate_2d_data_plot(data, data_raster, 
      interpolate = interpolate, title = title)
   return(plot_object)
}

If you modify it to this then it might work.

   if (class(model) != "ranger"){
   predictions <- predict(model, grid, ...)
   } else {
   predictions <- as.character(predict(model, grid, ...)$predictions)   
   }

If it can make a ggplot object then I think I can use GGally to make a grid like figure 1. I think this could be an interesting and useful diagnostic in a scatterplot matrix style plot, perhaps similar to GGally::ggpairs.