LogicForest (version 2.1.2)

logforest: Logic Forest & Logic Survival Forest

Description

Constructs an ensemble of logic regression models using bagging for classification or regression, and identifies important predictors and interactions. Logic Forest (LF) efficiently searches the space of logical combinations of binary variables using simulated annealing. It has been extended to support linear and survival regression.

Usage

logforest(
  resp.type,
  resp,
  resp.time = data.frame(X = rep(1, nrow(resp))),
  Xs,
  nBSXVars,
  anneal.params,
  nBS = 100,
  h = 0.5,
  norm = TRUE,
  numout = 5,
  nleaves
)

Value

A logforest object containing:

Predictor.frequency

Frequency of each predictor across trees.

Predictor.importance

Importance of each predictor.

PI.frequency

Frequency of each interaction across trees.

PI.importance

Importance of each interaction.

Arguments

resp.type

String indicating regression type: "bin" for classification, "lin" for linear regression, "exp_surv" for exponential time-to-event, and "cph_surv" for Cox proportional hazards.

resp

Numeric vector of response values (binary for classification/survival, continuous for linear regression). For time-to-event, indicates event/censoring status.

resp.time

Numeric vector of event/censoring times (used only for survival models).

Xs

Matrix or data frame of binary predictor variables (0/1 only).

nBSXVars

Integer. Number of predictors sampled for each tree (default is all predictors).

anneal.params

A list of parameters for simulated annealing (see logreg.anneal.control). Defaults: start = 1, end = -2, iter = 50000.

nBS

Number of trees to fit in the logic forest.

h

Numeric. Minimum proportion of trees predicting "1" required to classify an observation as "1" (used for classification).

norm

Logical. If FALSE, importance scores are not normalized.

numout

Integer. Number of predictors and interactions to report.

nleaves

Integer. Maximum number of leaves (end nodes) allowed per tree.

Author

Bethany J. Wolf
J. Madison Hyer

Details

Logic Forest is designed to identify interactions between binary predictors without requiring their pre-specification. Using simulated annealing, it searches the space of all possible logical combinations (e.g., AND, OR, NOT) among predictors. Originally developed for binary outcomes in gene-environment interaction studies, it has since been extended to linear and time-to-event outcomes (Logic Survival Forest).

References

Wolf BJ, Hill EG, Slate EH. (2010). Logic Forest: An ensemble classifier for discovering logical combinations of binary markers. Bioinformatics, 26(17):2183–2189. tools:::Rd_expr_doi("10.1093/bioinformatics/btq354")
Wolf BJ et al. (2012). LBoost: A boosting algorithm with application for epistasis discovery. PLoS One, 7(11):e47281. tools:::Rd_expr_doi("10.1371/journal.pone.0047281")
Hyer JM et al. (2019). Novel Machine Learning Approach to Identify Preoperative Risk Factors Associated With Super-Utilization of Medicare Expenditure Following Surgery. JAMA Surg, 154(11):1014–1021. tools:::Rd_expr_doi("10.1001/jamasurg.2019.2979")

See Also

pimp.import, logreg.anneal.control

Examples

if (FALSE) {
set.seed(10051988)
N_c <- 50
N_r <- 200
init <- as.data.frame(matrix(0, nrow = N_r, ncol = N_c))
colnames(init) <- paste0("X", 1:N_c)
for(n in 1:N_c){
  p <- runif(1, min = 0.2, max = 0.6)
  init[,n] <- rbinom(N_r, 1, p)
}

X3X4int <- as.numeric(init$X3 == init$X4)
X5X6int <- as.numeric(init$X5 == init$X6)
y_p <- -2.5 + init$X1 + init$X2 + 2 * X3X4int + 2 * X5X6int
p <- 1 / (1 + exp(-y_p))
init$Y.bin <- rbinom(N_r, 1, p)

# Classification
LF.fit.bin <- logforest("bin", init$Y.bin, NULL, init[,1:N_c], nBS=10, nleaves=8, numout=10)
print(LF.fit.bin)

# Continuous
init$Y.cont <- rnorm(N_r, mean = 0) + init$X1 + init$X2 + 5 * X3X4int + 5 * X5X6int
LF.fit.lin <- logforest("lin", init$Y.cont, NULL, init[,1:N_c], nBS=10, nleaves=8, numout=10)
print(LF.fit.lin)

# Time-to-event
shape <- 1 - 0.05*init$X1 - 0.05*init$X2 - 0.2*init$X3*init$X4 - 0.2*init$X5*init$X6
scale <- 1.5 - 0.05*init$X1 - 0.05*init$X2 - 0.2*init$X3*init$X4 - 0.2*init$X5*init$X6
init$TIME_Y <- rgamma(N_r, shape = shape, scale = scale)
LF.fit.surv <- logforest("exp_surv", init$Y.bin, init$TIME_Y, init[,1:N_c],
  nBS=10, nleaves=8, numout=10)
print(LF.fit.surv)
}