R neuralnet package: Can't train neural network

Question

I'm trying to use the neuralnet package to train a model on this data set. However, I'm getting the following error which I can't understand:

Error: the error derivative contains a NA; varify that the derivative function does not divide by 0 (e.g. cross entropy)

This is my code:

library(neuralnet)
library(tidyverse)

framingham <- read_csv('https://courses.edx.org/assets/courseware/v1/7022cf016eefb6d3747447589423dab0/asset-v1:MITx+15.071x+3T2019+type@asset+block/framingham.csv',
                       col_types = cols(.default = 'i',sysBP = 'n', diaBP = 'n', BMI = 'n' ))
# Split data
set.seed(123); train_idx <- sample(nrow(framingham), 2/3 * nrow(framingham))
framingham_train <- framingham[train_idx, ]
framingham_test <- framingham[-train_idx, ]

# Binary classification
nn <- neuralnet(formula = TenYearCHD ~ ., data = framingham_train,
                hidden=c(3,2),
                act.fct = "tanh",
                stepmax = 1e8,
                err.fct = 'ce',
                linear.output = TRUE)

I've tried changing the error function and other details, but nothing seem to work.

Answer 1

There are some columns with NAs:

summary(framingham)
      male             age          education     currentSmoker   
 Min.   :0.0000   Min.   :32.00   Min.   :1.000   Min.   :0.0000  
 1st Qu.:0.0000   1st Qu.:42.00   1st Qu.:1.000   1st Qu.:0.0000  
 Median :0.0000   Median :49.00   Median :2.000   Median :0.0000  
 Mean   :0.4292   Mean   :49.58   Mean   :1.979   Mean   :0.4941  
 3rd Qu.:1.0000   3rd Qu.:56.00   3rd Qu.:3.000   3rd Qu.:1.0000  
 Max.   :1.0000   Max.   :70.00   Max.   :4.000   Max.   :1.0000  
                                  NA's   :105                     
   cigsPerDay         BPMeds        prevalentStroke     prevalentHyp   
 Min.   : 0.000   Min.   :0.00000   Min.   :0.000000   Min.   :0.0000  
 1st Qu.: 0.000   1st Qu.:0.00000   1st Qu.:0.000000   1st Qu.:0.0000  
 Median : 0.000   Median :0.00000   Median :0.000000   Median :0.0000  
 Mean   : 9.006   Mean   :0.02962   Mean   :0.005896   Mean   :0.3106  
 3rd Qu.:20.000   3rd Qu.:0.00000   3rd Qu.:0.000000   3rd Qu.:1.0000  
 Max.   :70.000   Max.   :1.00000   Max.   :1.000000   Max.   :1.0000  
 NA's   :29       NA's   :53 

If you subset for rows with no NAs, it should work:

set.seed(123)
framingham = framingham[complete.cases(framingham),]
train_idx <- sample(nrow(framingham), 2/3 * nrow(framingham))
framingham_train <- framingham[train_idx, ]
framingham_test <- framingham[-train_idx, ]

Also, I think you cannot use tanh for activation with cross-entropy, so something below will work with logistic as activation function:

nn <- neuralnet(formula = TenYearCHD ~ ., act.fct="logistic",rep = 3,
data = framingham_train,hidden=c(3,2), err.fct = 'ce',linear.output=FALSE)

Answer 2

I also had this problem, but the right solution is to choose the right variables to enter the model. So you should not enter less important variables into the model