Laurae (version 0.0.0.9001)

CascadeForest: Cascade Forest implementation in R

Description

This function attempts to replicate Cascade Forest using xgboost. It performs Complete-Random Tree Forest in a Neural Network directed acrylic graph like in Neural Networks, but only for simple graphs (e.g use previous layer output data for next layer training each time). You can specify your learning objective using objective and the metric to check for using eval_metric. You can plug custom objectives instead of the objectives provided by xgboost. As with any uncalibrated machine learning methods, this method suffers uncalibrated outputs. Therefore, the usage of scale-dependent metrics is discouraged (please use scale-invariant metrics, such as Accuracy, AUC, R-squared, Spearman correlation...).

Usage

CascadeForest(training_data, validation_data, training_labels,
  validation_labels, folds, boosting = FALSE, nthread = 1, cascade_lr = 1,
  training_start = NULL, validation_start = NULL, cascade_forests = rep(4,
  5), cascade_trees = 500, cascade_rf = 2,
  cascade_seeds = 1:length(cascade_forests), objective = "reg:linear",
  eval_metric = Laurae::df_rmse, multi_class = FALSE, early_stopping = 2,
  maximize = FALSE, verbose = TRUE, low_memory = FALSE,
  essentials = FALSE, garbage = FALSE, work_dir = NULL,
  fail_safe = 65536)

Arguments

training_data

Type: data.table. The training data. Columns are added during training if low_memory == TRUE, so you may want to clean it up if you use low_memory == TRUE and interrupt training.

validation_data

Type: data.table. The validation data to check for metric performance. Set to NULL if you want to use out of fold validation data instead of a custom validation data set. Columns are added during training if low_memory == TRUE, so you may want to clean it up if you use low_memory == TRUE and interrupt training.

training_labels

Type: numeric vector. The training labels.

validation_labels

Type: numeric vector. The validation labels.

folds

Type: list. The folds as list for cross-validation.

boosting

Type: logical. Whether to perform boosting or not for training. It may converge faster, but may overfit faster and therefore needs control via cascade_lr. Defaults to FALSE.

nthread

Type: numeric. The number of threads using for multithreading. 1 means singlethread (uses only one core). Higher may mean faster training if the memory overhead is not too large. Defaults to 1.

cascade_lr

Type: numeric vector or numeric. The shrinkage affected to each tree per layer to avoid overfitting, for each layer. You may specify a vector to change the learning rate per layer, such as c(0.4, 0.3, 0.2, 0.1, 0.05) so you can perform boosting afterwards. Defaults to 1.

training_start

Type: numeric vector. The initial training prediction labels. Set to NULL if you do not know what you are doing. Defaults to NULL.

validation_start

Type: numeric vector. The initial validation prediction labels. Set to NULL if you do not know what you are doing. Defaults to NULL.

cascade_forests

Type: numeric vector (mandatory). The number of forest models per layer in the architecture to create for the Cascade Forest. Inputting 0 means it will take the latest number forever until it stops after convergence using early_stopping. For instance, to activate infinite training on the default parameters, use c(rep(4, 5), 0). Defaults to rep(4, 5).

cascade_trees

Type: numeric vector or numeric. The number of trees per forest model per layer in the architecture to create for the Cascade Forest. You may specify a vector to change the learning rate per layer, such as 500 so you can perform boosting afterwards. Defaults to 1000.

cascade_rf

Type: numeric vector or numeric. The number of Random Forest model per layer in the architecture to create for the Cascade Forest. You may specify a vector to change the learning rate per layer, such as c(1, 1, 2, 3, 5) so you can perform boosting afterwards. Defaults to 2.

cascade_seeds

Type: numeric vector or numeric. Random seed for reproducibility per layer. Do not set it to a value which is identical throughout the architecture, you will train on the same features over and over otherwise! When using a single value as seed, it automatically adds 1 each time an advance in the layer is made. Defaults to 1:length(cascade_forests).

objective

Type: character or function. The function which leads boosting loss. See xgboost::xgb.train. Defaults to "reg:linear".

eval_metric

Type: function. The function which evaluates boosting loss. Must take two arguments in the following order: preds, labels (they may be named in another way) and returns a metric. Defaults to Laurae::df_rmse.

multi_class

Type: numeric. Defines the number of classes internally for whether you are doing multi class classification or not to use specific routines for multiclass problems when using return_list == FALSE. Defaults to 2, which is for regression and binary classification.

early_stopping

Type: numeric. Defines how many architecture layers without improvement to require before stopping early (therefore, you must remove 1 to that value - for instance, a stopping of 2 means it will stop after 3 failures to improve). 0 means instantly stop at the first failure for improvement. -1 means no stopping. Requires validation_data to be able to stop early. Defaults to 2.

maximize

Type: logical. Whether to maximize or not the loss evaluation metric. Defaults to FALSE.

verbose

Type: logical. Whether to print training evaluation. Defaults to TRUE.

low_memory

Type: logical. Whether to perform the data.table transformations in place to lower memory usage. Defaults to FALSE.

essentials

Type: logical. Whether to store intermediary predictions or not. Set it to TRUE if you encounter memory issues. Defaults to FALSE.

garbage

Type: logical. Whether to perform garbage collect regularly. Defaults to FALSE.

work_dir

Type: character, allowing concatenation with another character text (ex: "dev/tools/save_in_this_folder/" = add slash, or "dev/tools/save_here/prefix_" = don't add slash). The working directory to store models. If you provide a working directory, the models will be saved inside that directory (and all other models will get wiped if they are under the same names). It will lower severely the memory usage as the models will not be saved anymore in memory. Combined with garbage == TRUE, you achieve the lowest possible memory usage in this Deep Forest implementation. Defaults to NULL, which means store models in memory.

fail_safe

Type: numeric. In case of infinite training (cascade_forests's last value equal to 0), this limits the number of training iterations. Defaults to 65536.

Value

A data.table based on target.

Details

For implementation details of Cascade Forest / Complete-Random Tree Forest / Multi-Grained Scanning / Deep Forest, check this: https://github.com/Microsoft/LightGBM/issues/331#issuecomment-283942390 by Laurae. Cascade Forests tend to aim what Neural Networks are doing: architecturing the model in multiple layers. Cascade Forests abuse the stacking ensemble method to perform training on these layers. Using randomness of Random Forests and Complete-Random Tree Forests, a Cascade Forest aims to outperform simple Convolutional Neural Networks (CNNs). The computational cost, however, is massive and should be taken into account before learning a large (and potentially unlimited) Cascade Forest. Due to their nature and to stacking ensemble properties, Cascade Forests have a hard time to overfit themselves. Putting a Cascade Forest on top a Multi-Grained Scanning model results in a gcForest Laurae recommends using xgboost or LightGBM on top of gcForest or Cascade Forest. See the rationale here: https://github.com/Microsoft/LightGBM/issues/331#issuecomment-284689795.

Examples

## Not run: ------------------------------------
# # Load libraries
# library(data.table)
# library(Matrix)
# library(xgboost)
# 
# # Create data
# data(agaricus.train, package = "lightgbm")
# data(agaricus.test, package = "lightgbm")
# agaricus_data_train <- data.table(as.matrix(agaricus.train$data))
# agaricus_data_test <- data.table(as.matrix(agaricus.test$data))
# agaricus_label_train <- agaricus.train$label
# agaricus_label_test <- agaricus.test$label
# folds <- Laurae::kfold(agaricus_label_train, 5)
# 
# # Train a model (binary classification)
# model <- CascadeForest(training_data = agaricus_data_train, # Training data
#                        validation_data = agaricus_data_test, # Validation data
#                        training_labels = agaricus_label_train, # Training labels
#                        validation_labels = agaricus_label_test, # Validation labels
#                        folds = folds, # Folds for cross-validation
#                        boosting = FALSE, # Do not touch this unless you are expert
#                        nthread = 1, # Change this to use more threads
#                        cascade_lr = 1, # Do not touch this unless you are expert
#                        training_start = NULL, # Do not touch this unless you are expert
#                        validation_start = NULL, # Do not touch this unless you are expert
#                        cascade_forests = rep(4, 5), # Number of forest models
#                        cascade_trees = 10, # Number of trees per forest
#                        cascade_rf = 2, # Number of Random Forest in models
#                        cascade_seeds = 1:5, # Seed per layer
#                        objective = "binary:logistic",
#                        eval_metric = Laurae::df_logloss,
#                        multi_class = 2, # Modify this for multiclass problems
#                        early_stopping = 2, # stop after 2 bad combos of forests
#                        maximize = FALSE, # not a maximization task
#                        verbose = TRUE, # print information during training
#                        low_memory = FALSE)
# 
# # Attempt to perform fake multiclass problem
# agaricus_label_train[1:100] <- 2
# 
# # Train a model (multiclass classification)
# model <- CascadeForest(training_data = agaricus_data_train, # Training data
#                        validation_data = agaricus_data_test, # Validation data
#                        training_labels = agaricus_label_train, # Training labels
#                        validation_labels = agaricus_label_test, # Validation labels
#                        folds = folds, # Folds for cross-validation
#                        boosting = FALSE, # Do not touch this unless you are expert
#                        nthread = 1, # Change this to use more threads
#                        cascade_lr = 1, # Do not touch this unless you are expert
#                        training_start = NULL, # Do not touch this unless you are expert
#                        validation_start = NULL, # Do not touch this unless you are expert
#                        cascade_forests = rep(4, 5), # Number of forest models
#                        cascade_trees = 10, # Number of trees per forest
#                        cascade_rf = 2, # Number of Random Forest in models
#                        cascade_seeds = 1:5, # Seed per layer
#                        objective = "multi:softprob",
#                        eval_metric = Laurae::df_logloss,
#                        multi_class = 3, # Modify this for multiclass problems
#                        early_stopping = 2, # stop after 2 bad combos of forests
#                        maximize = FALSE, # not a maximization task
#                        verbose = TRUE, # print information during training
#                        low_memory = FALSE)
## ---------------------------------------------