Statistics for Data Science

Training

Boosting is designed to work best with weak learners, so a very small number of hidden neurons in the model's hidden layer are used.

Concretely, we will begin with the simplest possible multilayer perceptron that uses a single hidden neuron. To understand the effect of using boosting, a baseline performance is established by training a single neural network (and measuring its performance).

This is to accomplish the following:

> library(nnet) 
> n_model <- nnet(CLASS ~ ., data = magic_train_df_pp, size = 1) 
> n_test_predictions <- predict(n_model, magic_test_pp, 
                                type = "class") 
> (n_test_accuracy <- mean(n_test_predictions ==   
                           magic_test_output)) 
[1] 0.7948988 

This establishes that we have a baseline accuracy of around 79.5 percent. ...