rrf.opt.m: KnowGRRF with weights from multiple knowledge domain

In KnowGRRF: Knowledge-Based Guided Regularized Random Forest

Description

Regularize on the weights to guide RRF feature selection. Weights can from multiple knowledge domain and/or combination with statistics-based weights, e.g., p/q value, variable importance, etc. Proportion of weights can be scaled by regularization parameters. Feature set selected is also based on stability, that is the frequency of selection from multiple runs. Features that are consistently selected from multiple runs will be used in a random forest model, from which AIC and AUC will be calculated to evaluate the model performance. Only AIC will be calculated for regression.

Usage

rrf.opt.m(X.train, Y.train, X.test=NULL, Y.test=NULL, pwr, 
weight, iter=1,total=10, cutoff=0.5)

Arguments

X.train	a data frame or matrix (like x) containing predictors for the training set.
Y.train	response for the training set. If a factor, classification is assumed, otherwise regression is assumed. If omitted, will run in unsupervised mode.
X.test	an optional data frame or matrix (like x) containing predictors for the test set.
Y.test	optional response for the test set.
pwr	Regularization term to adjust the scale of weights. When multiple domain knowledge is used, pwr is a vector, with length equal to the number of domain knowledge plus one. First parameter is the scaling parameter, and the rest of the vectors correspond to the relative importance of each domain knowledge. Larger regularization will differentiate the importance of variables more significantly. Fewer variables tend to be selected with large pwr. This parameter can be tuned using optimization methods or grid searching with on.aic function.
weight	A matrix of weights corresponding to each of predictors. Each column correspond to each domain knowledge and each row correspond to each variable. Weights are between 0 and 1.
iter	The number of RF model built to evaluate AIC and AUC. AIC is calculated using out-of-bag prediction from random forest using feature selected. AUC is calculated for classification problem only.
total	the number of times to repeat the selection for stability test in select.stable function.
cutoff	The minimum percentage of times that the feature is selected in multiple runs for stability test, ranges between 0 and 1.

Value

return a list, including

AIC	AIC calculated from random forest model out-of-bag predicted probability for classification, or out-of-bag prediction for classification
AUC	AUC calculated from out-of-bag prediction from random forest classification model
Test.AUC	AUC calculated from test prediction from random forest classification model
AUC	AUC calculated from out-of-bag prediction from random forest classification model
feaSet	feature set selected

Note

This function can be used after weights and regularization term are determined. Weights are from knowledege domain and regularization term can be determined by optimization. See example.

Author(s)

Xin Guan, Li Liu

References

Guan, X., & Liu, L. (2018). Know-GRRF: Domain-Knowledge Informed Biomarker Discovery with Random Forests.

Examples

##---- Example: regression ----
library(randomForest)

set.seed(1)
X.train<-data.frame(matrix(rnorm(100*100), nrow=100))
b=seq(1, 6, 0.5) 
##y has a linear relationship with first 10 variables
y.train=b[7]*X.train$X6+b[8]*X.train$X7+b[9]*X.train$X8+b[10]*X.train$X9+b[11]*X.train$X10 


##use weights from domain knowledge. If not available, 
##can use statistic-based weights, e.g., variable importance, p/q value, etc
prior1 <- abs(c(rnorm(5, 5, 1), rnorm(95, 0, 1)))  
##domain 1 suggest first five are important variables
prior2 <- abs(c(rnorm(5, 0, 1), rnorm(5, 8, 2), rnorm(90, 0, 1)))  
##domain 2 suggest next five are important variables
imp<-randomForest(X.train, y.train)$importance 
prior3=0.5+0.5*imp/max(imp)   ##domain 3 uses relative varialbe importance



#'\donttest{
#'use optimization function to find the appropriate regularization term 
#'to scale weights and then apply the weights to guide the RRF

#'opt<-optim(par=c(1,1,1,1), fn=on.aic, X.train=X.train, Y.train=y.train, 
#'weight=cbind(prior1, prior2, prior3), iter=5,total=10, cutoff=0.5, num = 3, 
#'method='L-BFGS-B', lower=0.01, upper=0.5, control=list(fnscale=1,trace = TRUE)) 
#'can take long for four parameters to be optimized.
#'opt$par can be used as input of pwr in rrf.opt.m
#'}

rrf.opt.m(X.train, y.train, pwr=c(5,1,1,1), weight=cbind(prior1, prior2, prior3))  

KnowGRRF documentation built on May 2, 2019, 6:43 a.m.