Runs the Mossy Forest algorithm.
Usage
mf(
X,
y,
module_membership,
screen_params = screen_control(),
select_params = select_control(),
final_ntree = 500L,
nodesize = NULL,
num_processors = 1L,
max_modsize = 5000L,
verbose = 1L,
seed = as.integer(Sys.time()),
compute_interactions = TRUE,
backend = "python",
r_shap = "permutation"
)Arguments
- X
A data frame where each column is a feature.
- y
Response vector (numeric for regression, factor for classification).
- module_membership
A named character vector assigning each feature in
Xto a module. Names must matchcolnames(X). Modules named inselect_control(unassigned_modules)form the unassigned pool, which is selected with its own elbow based threshold; all other modules usepi_thr.- screen_params
Screening parameters. See
screen_control.- select_params
Selection parameters. See
select_control.- final_ntree
Number of trees in the final random forest. Default
500.- nodesize
Minimum terminal node size. Defaults to
1for classification and5for regression.- num_processors
Number of parallel threads for random forests. Default
1.- max_modsize
Maximum module size before blockwise pre reduction is applied (Python backend only). Default
5000.- verbose
Verbose level.
0= silent;1= progress messages;2= detailed output. Default1.- seed
Integer random seed for reproducibility. Defaults to current system time.
- compute_interactions
Logical. If
TRUE(default), the TreeSHAP interaction matrix is computed by the Python backend and by the pure R backend withr_shap = "treeshap". SetFALSEto skip it (faster);detect_interaction()andplot_potential_interactions()then have nothing to use. It has no effect withr_shap = "permutation", which never computes interactions.- backend
Computation backend.
"python"(default) uses a Python scikit-learn / SHAP engine via reticulate and provides exact TreeSHAP values plus an interaction matrix."R"uses a R ranger implementation, no Python installation required. Choose the final importance method withr_shap.- r_shap
Final SHAP / importance method used when
backend = "R". Ignored for the Python backend. One of:"permutation"Ranger permutation variable importance (default). No extra packages needed. Returns a scalar VIM per feature; per observation SHAP values and SHAP based plots are unavailable.
"treeshap"Exact TreeSHAP values via the R treeshap package. Requires treeshap. Also computes a \(p \times p\) interaction matrix, enabling
detect_interaction()andplot_potential_interactions().
Value
An object of class mossy_forest.
References
Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W., & Robins, J. (2018). Double/cross-fit residualed machine learning for treatment and structural parameters. The Econometrics Journal, 21(1), C1-C68.
Lundberg, S. M., & Lee, S. I. (2017). A unified model for interpreting predictions. NeurIPS, 30.
Examples
# A small simulated regression problem with two modules of features
set.seed(1)
n <- 60
z <- rnorm(n)
X <- data.frame(a1 = z + rnorm(n, sd = 0.3), a2 = z + rnorm(n, sd = 0.3),
a3 = z + rnorm(n, sd = 0.3), b1 = rnorm(n), b2 = rnorm(n),
b3 = rnorm(n))
y <- z + rnorm(n, sd = 0.3)
membership <- c(a1 = "A", a2 = "A", a3 = "A", b1 = "B", b2 = "B", b3 = "B")
# The pure R backend needs no Python installation
fit <- mf(X, y, module_membership = membership,
screen_params = screen_control(min_ntree = 50),
select_params = select_control(n_reps = 10, min_ntree = 50),
final_ntree = 50, backend = "R", verbose = 0, seed = 1)
print(fit)
#> --- Mossy Forest ---
#> Stable features selected: 2
#>
#> feature_name variable_importance module_membership
#> 1 a1 0.5034 A
#> 2 a2 0.4403 A
#>
#> OOB prediction error: 0.1831
head(fit$final_SHAP)
#> feature_name variable_importance module_membership
#> 1 a1 0.5034 A
#> 2 a2 0.4403 A
if (FALSE) { # \dontrun{
# The Python backend (exact TreeSHAP) needs a configured Python environment
mf_setup(condaenv = "sfenv")
fit_py <- mf(X, y, module_membership = membership, backend = "python")
plot_importance(fit_py)
} # }