Regularized Greedy Forest in R

14 Feb 2018

This blog post is about my newly released RGF package (the blog post consists mainly of the package Vignette). The RGF package is a wrapper of the Regularized Greedy Forest python package, which also includes a Multi-core implementation (FastRGF). Portability from Python to R was made possible using the reticulate package and the installation requires basic knowledge of Python. Except for the Linux Operating System, the installation on Macintosh and Windows might be somehow cumbersome (on windows the package currently can be used only from within the command prompt). Detailed installation instructions for all three Operating Systems can be found in the README.md file and in the rgf_python Github repository.

UPDATE 26-07-2018: A Singularity image file is available in case that someone intends to run RGF on Ubuntu Linux (locally or in a cloud instance) with all package requirements pre-installed. This allows the user to utilize the RGF package without having to spend time on the installation process.

The Regularized Greedy Forest algorithm is explained in detail in the paper Rie Johnson and Tong Zhang, Learning Nonlinear Functions Using Regularized Greedy Forest. A small synopsis would be “… the resulting method, which we refer to as regularized greedy forest (RGF), integrates two ideas: one is to include tree-structured regularization into the learning formulation; and the other is to employ the fully-corrective regularized greedy algorithm ….”.

At the time of writing this blog post (14 - 02 - 2018), there isn’t a corresponding implementation of the algorithm in the R language, so I decided to port the Python package in R taking advantage of the reticulate package. In the next lines, I will explain the functionality of the package and I compare RGF with other similar implementations, such as ranger (random forest algorithm) and xgboost (gradient boosting algorithm), in terms of time efficiency and error rate improvement.

The RGF package

The RGF package includes the following R6-classes / functions,

classes

RGF_Regressor	RGF_Classifier	FastRGF_Regressor	FastRGF_Classifier
fit()	fit(()	fit()	fit()
predict()	predict()	predict()	predict()
cleanup()	predict_proba()	cleanup()	predict_proba()
get_params()	cleanup()	get_params()	cleanup()
score()	get_params()	score()	get_params()
feature_importances()	score()		score()
dump_model()	feature_importances()
	dump_model()

functions

UPDATE 10-05-2018 : Beginning from version 1.0.3 the dgCMatrix_2scipy_sparse function was renamed to TO_scipy_sparse and now accepts either a dgCMatrix or a dgRMatrix as input. The appropriate format for the RGF package in case of sparse matrices is the dgCMatrix format (scipy.sparse.csc_matrix)

TO_scipy_sparse()

RGF_cleanup_temp_files()

mat_2scipy_sparse()

The package documentation includes details and examples for all R6-classes and functions. In the following code chunks, I’ll explain how a user can work with sparse matrices as all RGF algorithms (besides a dense matrix) require a python sparse matrix as input.

Sparse matrices as input

The RGF package includes two functions (mat_2scipy_sparse and TO_scipy_sparse) which allow the user to convert from a matrix / sparse matrix (dgCMatrix, dgRMatrix) to a scipy sparse matrix (scipy.sparse.csc_matrix, scipy.sparse.csr_matrix),

library(nmslibR)

# conversion from a matrix object to a scipy sparse matrix
#----------------------------------------------------------

set.seed(1)

x = matrix(runif(1000), nrow = 100, ncol = 10)

x_sparse = mat_2scipy_sparse(x, format = "sparse_row_matrix")

print(dim(x))

[1] 100  10

print(x_sparse$shape)

(100, 10)

# conversion from a dgCMatrix object to a scipy sparse matrix
#-------------------------------------------------------------

data = c(1, 0, 2, 0, 0, 3, 4, 5, 6)

# 'dgCMatrix' sparse matrix
#--------------------------

dgcM = Matrix::Matrix(data = data, nrow = 3,

                      ncol = 3, byrow = TRUE,

                      sparse = TRUE)

print(dim(dgcM))

[1] 3 3

x_sparse = TO_scipy_sparse(dgcM)

print(x_sparse$shape)

(3, 3)

# 'dgRMatrix' sparse matrix
#--------------------------

dgrM = as(dgcM, "RsparseMatrix")

class(dgrM)

# [1] "dgRMatrix"
# attr(,"package")
# [1] "Matrix"

print(dim(dgrM))

[1] 3 3

res_dgr = TO_scipy_sparse(dgrM)

print(res_dgr$shape)

(3, 3)

Comparison of RGF with ranger and xgboost

First the data, libraries and cross-validation function will be inputted (the MLmetrics library is also required),

data(Boston, package = 'KernelKnn')

library(RGF)
library(ranger)
library(xgboost)

# shuffling function for cross-validation folds
#-----------------------------------------------

func_shuffle = function(vec, times = 10) {

  for (i in 1:times) {

    out = sample(vec, length(vec))
  }
  out
}

# cross-validation folds [ regression]
#-------------------------------------

regr_folds = function(folds, RESP, stratified = FALSE) {

  if (is.factor(RESP)) {

    stop(simpleError("this function is meant for regression for classification use the 'class_folds' function"))
  }

  samp_vec = rep(1/folds, folds)

  sort_names = paste0('fold_', 1:folds)

  if (stratified == TRUE) {

    stratif = cut(RESP, breaks = folds)

    clas = lapply(unique(stratif), function(x) which(stratif == x))

    len = lapply(clas, function(x) length(x))

    prop = lapply(len, function(y) sapply(1:length(samp_vec), function(x) round(y * samp_vec[x])))

    repl = unlist(lapply(prop, function(x) sapply(1:length(x), function(y) rep(paste0('fold_', y), x[y]))))

    spl = suppressWarnings(split(1:length(RESP), repl))}

  else {

    prop = lapply(length(RESP), function(y) sapply(1:length(samp_vec), function(x) round(y * samp_vec[x])))

    repl = func_shuffle(unlist(lapply(prop, function(x) sapply(1:length(x), function(y) rep(paste0('fold_', y), x[y])))))

    spl = suppressWarnings(split(1:length(RESP), repl))
  }

  spl = spl[sort_names]

  if (length(table(unlist(lapply(spl, function(x) length(x))))) > 1) {

    warning('the folds are not equally split')
  }

  if (length(unlist(spl)) != length(RESP)) {

    stop(simpleError("the length of the splits are not equal with the length of the response"))
  }

  spl
}

single threaded ( small data set )

In the next code chunk, I’ll perform 5-fold cross-validation using the Boston dataset and I’ll compare time execution and error rate for all three algorithms (comparison without doing hyper-parameter tuning),

NUM_FOLDS = 5

set.seed(1)
FOLDS = regr_folds(folds = NUM_FOLDS, Boston[, 'medv'], stratified = T)

boston_rgf_te = boston_ranger_te = boston_xgb_te = boston_rgf_time = boston_ranger_time = boston_xgb_time = rep(NA, NUM_FOLDS)

for (i in 1:length(FOLDS)) {

  cat("fold : ", i, "\n")

  samp = unlist(FOLDS[-i])
  samp_ = unlist(FOLDS[i])

  # RGF
  #----

  rgf_start = Sys.time()

  init_regr = RGF_Regressor$new(l2 = 0.1)

  init_regr$fit(x = as.matrix(Boston[samp, -ncol(Boston)]), y = Boston[samp, ncol(Boston)])

  pr_te = init_regr$predict(as.matrix(Boston[samp_, -ncol(Boston)]))

  rgf_end = Sys.time()

  boston_rgf_time[i] = rgf_end - rgf_start

  boston_rgf_te[i] = MLmetrics::RMSE(Boston[samp_, 'medv'], pr_te)

  # ranger
  #-------

  ranger_start = Sys.time()

  fit = ranger(dependent.variable.name = "medv", data = Boston[samp, ], write.forest = TRUE, 
               
               probability = F, num.threads = 1, num.trees = 500, verbose = T, 
               
               classification = F, mtry = NULL, min.node.size = 5, keep.inbag = T)

  pred_te = predict(fit, data = Boston[samp_, -ncol(Boston)], type = 'se')$predictions

  ranger_end = Sys.time()

  boston_ranger_time[i] = ranger_end - ranger_start

  boston_ranger_te[i] = MLmetrics::RMSE(Boston[samp_, 'medv'], pred_te)

  # xgboost
  #--------

  xgb_start = Sys.time()

  dtrain <- xgb.DMatrix(data = as.matrix(Boston[samp, -ncol(Boston)]), label = Boston[samp, ncol(Boston)])

  dtest <- xgb.DMatrix(data = as.matrix(Boston[samp_, -ncol(Boston)]), label = Boston[samp_, ncol(Boston)])

  watchlist <- list(train = dtrain, test = dtest)

  param = list("objective" = "reg:linear", "bst:eta" = 0.05, "max_depth" = 4, 
               
               "subsample" = 0.85, "colsample_bytree" = 0.85, "booster" = "gbtree",
               
               "nthread" = 1)

  fit = xgb.train(param, dtrain, nround = 500, print_every_n  = 100, watchlist = watchlist, early_stopping_rounds = 20,
                  
                  maximize = FALSE, verbose = 0)

  p_te = xgboost:::predict.xgb.Booster(fit, as.matrix(Boston[samp_, -ncol(Boston)]), ntreelimit = fit$best_iteration)

  xgb_end = Sys.time()

  boston_xgb_time[i] = xgb_end - xgb_start

  boston_xgb_te[i] = MLmetrics::RMSE(Boston[samp_, 'medv'], p_te)
}

fold :  1 
fold :  2 
fold :  3 
fold :  4 
fold :  5 

cat("total time rgf 5 fold cross-validation : ", sum(boston_rgf_time), " mean rmse on test data : ", mean(boston_rgf_te), "\n")

cat("total time ranger 5 fold cross-validation : ", sum(boston_ranger_time), " mean rmse on test data : ", mean(boston_ranger_te), "\n")

cat("total time xgb 5 fold cross-validation : ", sum(boston_xgb_time), " mean rmse on test data : ", mean(boston_xgb_te), "\n")

total time rgf 5 fold cross-validation :  0.7730639  mean rmse on test data :  3.832135 
total time ranger 5 fold cross-validation :  3.826846  mean rmse on test data :  4.17419 
total time xgb 5 fold cross-validation :  0.4316094  mean rmse on test data :  3.949122 

5 threads ( high dimensional dataset and presence of multicollinearity )

For the high-dimensional data (can be downloaded from my Github repository) I’ll use the FastRGF_Regressor rather than the RGF_Regressor (comparison without doing hyper-parameter tuning),

# download the data from my Github repository (tested on a Linux OS)

system("wget https://raw.githubusercontent.com/mlampros/DataSets/master/africa_soil_train_data.zip")

# load the data in the R session

train_dat = read.table(unz("africa_soil_train_data.zip", "train.csv"), nrows = 1157, header = T, quote = "\"", sep = ",")

# c("Ca", "P", "pH", "SOC", "Sand") : response variables            

# exclude response-variables and factor variable

x = train_dat[, -c(1, which(colnames(train_dat) %in% c("Ca", "P", "pH", "SOC", "Sand", "Depth")))]

# take (randomly) the first of the responses for train

y = train_dat[, "Ca"]

# dataset for ranger

tmp_rg_dat = cbind(Ca = y, x)

# cross-validation folds

set.seed(2)
FOLDS = regr_folds(folds = NUM_FOLDS, y, stratified = T)

highdim_rgf_te = highdim_ranger_te = highdim_xgb_te = highdim_rgf_time = highdim_ranger_time = highdim_xgb_time = rep(NA, NUM_FOLDS)

for (i in 1:length(FOLDS)) {

  cat("fold : ", i, "\n")

  new_samp = unlist(FOLDS[-i])
  new_samp_ = unlist(FOLDS[i])

  # RGF
  #----

  rgf_start = Sys.time()

  init_regr = FastRGF_Regressor$new(n_jobs = 5, l2 = 0.1)                  # I added 'l2' regularization

  init_regr$fit(x = as.matrix(x[new_samp, ]), y = y[new_samp])

  pr_te = init_regr$predict(as.matrix(x[new_samp_, ]))

  rgf_end = Sys.time()

  highdim_rgf_time[i] = rgf_end - rgf_start

  highdim_rgf_te[i] = MLmetrics::RMSE(y[new_samp_], pr_te)

  # ranger
  #-------

  ranger_start = Sys.time()
  
  fit = ranger(dependent.variable.name = "Ca", data = tmp_rg_dat[new_samp, ], 
               
               write.forest = TRUE, probability = F, num.threads = 5, num.trees = 500,
               
               verbose = T, classification = F, mtry = NULL, min.node.size = 5, 
               
               keep.inbag = T)
  
  pred_te = predict(fit, data = x[new_samp_, ], type = 'se')$predictions

  ranger_end = Sys.time()

  highdim_ranger_time[i] = ranger_end - ranger_start

  highdim_ranger_te[i] = MLmetrics::RMSE(y[new_samp_], pred_te)

  # xgboost
  #--------

  xgb_start = Sys.time()

  dtrain <- xgb.DMatrix(data = as.matrix(x[new_samp, ]), label = y[new_samp])

  dtest <- xgb.DMatrix(data = as.matrix(x[new_samp_, ]), label = y[new_samp_])

  watchlist <- list(train = dtrain, test = dtest)

  param = list("objective" = "reg:linear", "bst:eta" = 0.05, "max_depth" = 6, 
               
               "subsample" = 0.85, "colsample_bytree" = 0.85, "booster" = "gbtree",
               
               "nthread" = 5)                                                                     # "lambda" = 0.1 does not improve RMSE

  fit = xgb.train(param, dtrain, nround = 500, print_every_n  = 100, watchlist = watchlist,

                  early_stopping_rounds = 20, maximize = FALSE, verbose = 0)

  p_te = xgboost:::predict.xgb.Booster(fit, as.matrix(x[new_samp_, ]), ntreelimit = fit$best_iteration)

  xgb_end = Sys.time()

  highdim_xgb_time[i] = xgb_end - xgb_start

  highdim_xgb_te[i] = MLmetrics::RMSE(y[new_samp_], p_te)
}

fold :  1 
fold :  2 
fold :  3 
fold :  4 
fold :  5 

cat("total time rgf 5 fold cross-validation : ", sum(highdim_rgf_time), " mean rmse on test data : ", mean(highdim_rgf_te), "\n")

cat("total time ranger 5 fold cross-validation : ", sum(highdim_ranger_time), " mean rmse on test data : ", mean(highdim_ranger_te), "\n")

cat("total time xgb 5 fold cross-validation : ", sum(highdim_xgb_time), " mean rmse on test data : ", mean(highdim_xgb_te), "\n")

total time rgf 5 fold cross-validation :  92.31971  mean rmse on test data :  0.5155166
total time ranger 5 fold cross-validation :  27.32866  mean rmse on test data :  0.5394164
total time xgb 5 fold cross-validation :  30.48834  mean rmse on test data :  0.5453544

The README.md file of the RGF package includes the SystemRequirements and installation instructions.

An updated version of the RGF package can be found in my Github repository and to report bugs/issues please use the following link, https://github.com/mlampros/RGF/issues.

mlampros Organizing and Sharing thoughts, Receiving constructive feedback