In andrew-mcinerney/statnn: Interpreting Neural Networks From a Statistical Perspective

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

To begin, load the interpretnn package.

library(interpretnn)

Also, load the package that we will use to fit the neural network. Our package works with a number of popular R packages for neural networks, and here we will use the nnet package.

library(nnet)

Now, load the data.

# load data ---------------------------------------------------------------
data(Boston)

Next, we fit a neural network. We will fit a neural network on the respose variable, medv, using all covariates and with two hidden nodes. As neural networks require random initial weights to begin learning, we use set.seed() for reproducibility.

set.seed(100)
nn <- nnet(medv ~ ., data = Boston, size = 2, trace = FALSE,
           linout = TRUE, maxit = 1000)

We can then create a interpretnn object

intnn <- interpretnn(nn, X = Boston[, -ncol(Boston)]) 

A useful summary table can then be produced using the summary() function

summary(intnn)

This tells provides us with simple point estimates of the effects, and the results from the multiple-parameter Wald test for each input.

We can visualise the covariate effects and their associated uncertainty using the plot() function, which creates Partial Covariate Effect (PCE) plots.

There is also a plotnn() function that visualise the significance of each weight from the single-parameter Wald test.

plotnn(intnn)

andrew-mcinerney/statnn documentation built on June 30, 2024, 4:09 p.m.