README.md
In bert9bert/ParallelForest: Random Forest Classification with Parallel Computing

ParallelForest R Package

ParallelForest is an R package implementing random forest classification using parallel computing, built with Fortran and OpenMP in the backend.

Install ParallelForest by entering the following into an R console.

install.packages("devtools")
devtools::install_github("bert9bert/ParallelForest")
library(ParallelForest)

This will use the latest source files maintained on GitHub to compile and install the package on the user's side. Alternative installation methods using pre-compiled binary packages are kept in the prebuilt_packages/ directory. The ParallelForest package is also available on CRAN, but the version on GitHub (here) may be more current.

The grow.forest function is used to grow (fit) a random forest, which returns a fitted forest object. Afterwards, the predict function can be used to make predictions on the fitted forest with new data. The following commands can be used to view documentation for these functions.

?grow.forest             # get the help page for growing a forest
?predict.forest          # get the help page for predicting on a fitted forest
?accessors.forest        # get help page for accessing elements of a fitted forest,
                         #     e.g. fittedforest["min_node_obs"]

A dataset of income and other person-level characteristics based off the U.S. Census Bureau's Current Population Surveys in 1994 and 1995 can be downloaded from the UCI Machine Learning Repository here. The dependent variable is binary and indicates whether the survey-taker's income is under or over a certain level. Seven predictor variables are used.

First, load the package into R, then load the training and testing datasets.

library(ParallelForest)

data(low_high_earners)       # cleaned and prepared training dataset (199,522 observations)
data(low_high_earners_test)  # cleaned and prepared testing dataset (99,761 observations)

Then fit a forest to the training data. A variety of hyper-parameters can be specified to control how the forest will be fit, which are documented in the help file (see ?grow.forest in R). If a hyper-parameter is not specified then the package may try to choose reasonable values for it.

fforest = grow.forest(
  Y~.,
  data=low_high_earners,
  min_node_obs = 1000,      # min obs to split node
  max_depth = 15,           # max tree depth
  numvars = 3,              # num vars to draw for each tree
  numboots = 50             # number of trees in the forest
)

Then use the fitted forest to get predictions on the training data and testing data. The predict function takes an argument called type to specify whether the returned predictions should be continuous probability values between 0 and 1 (by setting type="response") or should have only binary values 0 and 1 based off some threshold (by setting type="binary_0_1"). See the help file via ?predict.forest in R for more details.

# get probability predictions on the training data
fforest_train_pred = predict(fforest, low_high_earners, type="response")

# get probability predictions on the testing (hold-out) data
fforest_test_pred = predict(fforest, low_high_earners_test, type="response")

Compute the area under the ROC curve (AUC) for both the predictions on the training data and testing data to evaluate the fitted model's performance.

library(pROC)  # install the pROC package with `install.packages("pROC")` if needed

fforest_train_roc = pROC::roc(response=low_high_earners$Y, predictor=fforest_train_pred)
fforest_test_roc = pROC::roc(response=low_high_earners_test$Y, predictor=fforest_test_pred)

print(paste0("AUC on the training data is ", round(fforest_train_roc$auc, digits=3)*100, "%."))
print(paste0("AUC on the testing (hold-out) data is ", round(fforest_test_roc$auc, digits=3)*100, "%."))