ff.formula {fuzzyforest}

Fuzzy forests algorithm

Description

Implements formula interface for ff.

Usage

## S3 method for class 'formula'
ff(formula, data = NULL, module_membership, ...)

Arguments

formula

Formula object.

data

data used in the analysis.

module_membership

A character vector giving the module membership of each feature.

...

Additional arguments

Value

An object of type fuzzy_forest. This object is a list containing useful output of fuzzy forests. In particular it contains a data.frame with list of selected features. It also includes the random forest fit using the selected features.

Note

See ff for additional arguments. Note that the matrix, Z, of features that do not go through the screening step must specified separately from the formula. test_features and test_y are not supported in formula interface. As in the randomForest package, for large data sets the formula interface may be substantially slower.

This work was partially funded by NSF IIS 1251151 and AMFAR 8721SC.

References

Conn, D., Ngun, T., Ramirez C.M., Li, G. (2019). "Fuzzy Forests: Extending Random Forest Feature Selection for Correlated, High-Dimensional Data." Journal of Statistical Software, 91(9). doi: 10.18637/jss.v091.i09

Breiman, L. (2001). "Random Forests." Machine Learning, 45(1), 5-32. doi: 10.1023/A:1010933404324

Zhang, B. and Horvath, S. (2005). "A General Framework for Weighted Gene Co-Expression Network Analysis." Statistical Applications in Genetics and Molecular Biology, 4(1). doi: 10.2202/1544-6115.1128

See Also

ff, print.fuzzy_forest, predict.fuzzy_forest, modplot

Examples

#ff requires that the partition of the covariates be previously determined.
#ff is also handy if the user wants to test out multiple settings of WGCNA
#prior to running fuzzy forests.
library(mvtnorm)
gen_mod <- function(n, p, corr) {
  sigma <- matrix(corr, nrow=p, ncol=p)
  diag(sigma) <- 1
  X <- rmvnorm(n, sigma=sigma)
  return(X)
}

gen_X <- function(n, mod_sizes, corr){
  m <- length(mod_sizes)
  X_list <- vector("list", length = m)
  for(i in 1:m){
    X_list[[i]] <- gen_mod(n, mod_sizes[i], corr[i])
  }
  X <- do.call("cbind", X_list)
  return(X)
}

err_sd <- .5
n <- 500
mod_sizes <- rep(25, 4)
corr <- rep(.8, 4)
X <- gen_X(n, mod_sizes, corr)
beta <- rep(0, 100)
beta[c(1:4, 76:79)] <- 5
y <- X%*%beta + rnorm(n, sd=err_sd)
X <- as.data.frame(X)
dat <- as.data.frame(cbind(y, X))

Xtest <- gen_X(n, mod_sizes, corr)
ytest <- Xtest%*%beta + rnorm(n, sd=err_sd)
Xtest <- as.data.frame(Xtest)

cdist <- as.dist(1 - cor(X))
hclust_fit <- hclust(cdist, method="ward.D")
groups <- cutree(hclust_fit, k=4)

screen_c <- screen_control(keep_fraction = .25,
                           ntree_factor = 1,
                           min_ntree = 250)
select_c <- select_control(number_selected = 10,
                           ntree_factor = 1,
                           min_ntree = 250)

ff_fit <- ff(y ~ ., data=dat,
             module_membership = groups,
             screen_params = screen_c,
             select_params = select_c,
             final_ntree = 250)
#extract variable importance rankings
vims <- ff_fit$feature_list

#plot results
modplot(ff_fit)

#obtain predicted values for a new test set
preds <- predict(ff_fit, new_data=Xtest)

#estimate test set error
test_err <- sqrt(sum((ytest - preds)^2)/n)