preCluster: Extract decisions from a model and create bootstrap...

In aruaud/endoR: Interpret and Visualize Stable Tree Ensemble Models

Extract decisions from a model and create bootstrap resamples.

Description

to run before bootstrapping on parallel with the clustermq package and model2DE_cluster function. Extracts decisions, optionally discretizes them. Creates data partitions for bootstrapping.

Usage

preCluster(
  model,
  model_type,
  data,
  target,
  times = 10,
  p = 0.5,
  sample_weight = NULL,
  classPos = NULL,
  ntree = "all",
  maxdepth = Inf,
  dummy_var = NULL,
  discretize = FALSE,
  K = 2,
  mode = "data",
  seed = 0,
  in_parallel = FALSE,
  n_cores = detectCores() - 1
)

Arguments

model	model to extract rules from.
model_type	character string: 'RF', 'random forest', 'rf', 'xgboost', 'XGBOOST', 'xgb', 'XGB', 'ranger', 'Ranger', 'gbm' or 'GBM'.
data	data with the same columns than data used to fit the model.
target	response variable.
times	number of bootstraps
p	fraction of data to resample.
sample_weight	numeric vector with the weights of samples for bootstrap resampling. For classification, if 2 values are given, the 1st one is assumed to be for the positive class (classpos argument).
classPos	the positive class predicted by decisions
ntree	number of trees to use from the model (default = all)
maxdepth	maximal node depth to use for extracting rules (by default, full branches are used).
dummy_var	if multiclass variables were transformed into dummy variables before fitting the model, one can pass their names in a vector here to avoid multiple levels to be used in a same rule (recommended).
discretize	if TRUE, discretization is performed with K categories (discretizeDecisions, by default = FALSE).
K	numeric, number of categories to create from numeric variables (default: K = 2).
mode	whether to discretize variables based on the data distribution (default, mode = 'data') or on the data splits in the model (mode = 'model').
seed	which seed to use to make the random bootstraps - it is fixed for reproducibility
in_parallel	if TRUE, the function is run in parallel.
n_cores	if in_parallel = TRUE, and no cluster has been passed: number of cores to use.

Value

A list with the row numbers of partitioned data, the rules originally extracted from the model and new data if discretization was performed.

Examples

library(randomForest)
library(caret)

# import data and fit model
data(iris)
mod <- randomForest(Species ~ ., data = iris)

# Get decision ensemble with bootstrapping.

# Run 1 bootstrap after the other (times = 2 bootstraps)
endo_setosa <- model2DE_resampling(model = mod, model_type = "rf"
    , data = iris[, -5], target = iris$Species, classPos = "setosa"
    , times = 2, in_parallel = TRUE, n_cores = 2, filter = FALSE)

# Same but use different sample weights for bootstrapping
n_setosa <- sum(iris$Species == "setosa")
n_samp <- length(iris$Species)
samp_weight <- round(
    ifelse(iris$Species == "setosa", 1 - n_setosa/n_samp, n_setosa/n_samp)
    , digits = 2)
endo_setosa <- model2DE_resampling(model = mod, model_type = "rf"
    , data = iris[, -5], target = iris$Species, classPos = "setosa"
    , times = 2, sample_weight = samp_weight
    , in_parallel = TRUE, n_cores = 2, filter = FALSE)

# Run the bootstraps in parallel
# First do all steps before bootstrapping
preclu <- preCluster(model = mod, model_type = "rf", data = iris[, -5]
    , target = iris$Species, classPos = "setosa", times = 2
    , discretize = TRUE, in_parallel = FALSE)

# Remove the special characters from column names
colnames(preclu$data) <- compatibleNames(colnames(preclu$data))

# Parameters for clustermq: can also run on HPC environment
library(clustermq)
options(clustermq.scheduler = "multiprocess")
# ... and run in parallel on each bootstrap
# (preclu$partitions = list of sample indexes for each bootstraps)
endo_setosa <- Q(model2DE_cluster
    , partition = preclu$partitions
    , export = list(data = preclu$data
                , target = iris$Species
                , exec = preclu$exec
                , classPos = "setosa"
                , prune = TRUE, filter = FALSE
                , maxDecay = 0.05 # values needed for maxDecay and typeDecay
                , typeDecay = 2 # here default ones, see pruneDecisions()
                , in_parallel = FALSE # can parallelize within each boostrap!
           )
    , n_jobs = 2 # max number of bootstraps that can be ran in parallel
    , pkgs = c("data.table", "parallel", "caret", "stringr", "scales"
                , "dplyr", "inTrees", "endoR")
    , log_worker = FALSE # to keep a log of the runs, e.g. if it fails..
)


# Stability selection
# First we can look at the effect of the alpha parameter on selection;
# alpha = expected number of false decisions
alphas <- evaluateAlpha(rules = endo_setosa, alphas = c(1:5, 7, 10)
                        , data = preclu$data)
alphas$summary_table

# perform stability selection with alpha = 1
de_final <- stabilitySelection(rules = endo_setosa, alpha_error = 7)

# Plot the decision ensemble:
# Plants from the setosa species have small petal and narrow long sepals.
plotFeatures(de_final, levels_order = c("Low", "Medium", "High"))

# there is no interaction between variables (all decisions with len = 1,
# the number of variables in the rules)
de_final$rules_summary
# hence the network would be empty and couldn't be plotted...
# plotNetwork(de_final, hide_isolated_nodes = FALSE)

aruaud/endoR documentation built on Jan. 25, 2025, 2:20 a.m.