How to automate variable selection in glmnet and cross validation
I am learning the use of glmnet and brnn packages. Consider the following code:
library(RODBC)
library(brnn)
library(glmnet)
memory.limit(size = 4000)
z <-odbcConnect("mydb") # database with Access queries and tables
# import the data
f5 <- sqlFetch(z,"my_qry")
# head(f5)
# check for 'NA'
sum(is.na(f5))
# choose a 'locn', up to 16 of variable 'locn' are present
f6 <- subset(f5, locn == "mm")
# dim(f6)
# use glmnet to identify possible iv's
training_xnm <- f6[,1:52] # training data
xnm <- as.matrix(training_xnm)
y <- f6[,54] # response
fit.nm <- glmnet(xnm,y, family="binomial", alpha=0.6, nlambda=1000,standardize=TRUE,maxit=100000)
# print(fit.nm)
# cross validation for glmnet to determine a good lambda value
cv.fit.nm <- cv.glmnet(xnm, y)
# have a look at the 'min' and '1se' lambda values
cv.fit.nm$lambda.min
cv.fit.nm$lambda.1se
# returned $lambda.min of 0.002906279, $lambda.1se of 2.587214
# for testing purposes I choose a value between 'min' and '1se'
mid.lambda.nm = (cv.fit.nm$lambda.min + cv.fit.nm$lambda.1se)/2
print(coef(fit.nm, s = mid.lambda.nm)) # 8 iv's retained
# I then manually inspect the data frame and enter the column index for each of the iv's
# these iv's will be the input to my 'brnn' neural nets
cols <- c(1, 3, 6, 8, 11, 20, 25, 38) # column indices of useful iv's
# brnn creation: only one shown but this step will be repeated
# take a 85% sample from data frame
ridxs <- sample(1:nrow(f6), floor(0.85*nrow(f6)) ) # row id's
f6train <- f6[ridxs,] # the resultant data frame of 85%
f6train <-f6train[,cols] # 'cols' as chosen above
# For the 'brnn' phase response is a binary value, 'fin'
# and predictors are the 8 iv's found earlier
out = brnn( fin ~ x1 + x2 + x3 + x4 + x5 + x6 + x7 + x8, data=f6train, neurons=3,normalize=TRUE, epochs=500, verbose=FALSE)
#summary(out)
# see how well the net predicts the training cases
pred <- predict(out)
The above script runs OK.
My question is: How can I automate the above script to run for different values of locn, that is essentially how can I generalize getting the step: cols <- c(1, 3, 6, 8, 11, 20, 25, 38) # column indices of useful iv's. At present I can do this manually but cannot see how to do this in a general way for different values of locn, for example
locn.list <- c("am", "bm", "cm", "dm", "em")
for(j in 1:5) {
this.locn <- locn.list[j]
# run the above script
}
Answer
Since posting my question I have found a paper by Simon, Friedman, Hastie and Tibshirani: Coxnet: Regularized Cox Regression which addresses how to extract what I wanted.
Some relevant details from this paper and adapted for my data (except for symbol for lambda!): We can check which covariates our model chose to be active, and see the coefficients of those covariates.
coef(fit.nm, s = cv.fit.nm$lambda.min) # returns the p length coefficient vector
of the solution corresponding to lambda =cv.fit$lambda.min.
Coefficients <- coef(fit.nm, s = cv.fit.nm$lambda.min)
Active.Index <- which(Coefficients != 0)
Active.Coefficients <- Coefficients[Active.Index]
Active.Index # identifies the covariates that are active in the model and
Active.Coefficients # shows the coefficients of those covariates
Hope this may be of use to others!