Workshop Hands-on Machine Learning in R, October - November 2023 edition

by Katrien Antonio, Jonas Crevecoeur and Roel Henckaerts

Course materials for the Hands-on Machine Learning in R course.

📆 Module 1 on October 24 & 31, Module 2 on November 7 & 14 and Module 3 on November 21 & 28, 2023
🕧 From 9.30 am to 12.30 pm
📍 online, organized by Actuarieel Instituut

Course materials will be posted in the week before the workshop. You are now on the landing page for Module 3: neural networks.

Prework

The workshop requires a basic understanding of R. A good starting level is the material covered in the werkt-U-al-met-R? repo on Katrien’s GitHub.

Familiarity with statistical or machine learning methods is not required. The workshop gradually builds up these concepts, with an emphasis on hands-on demonstrations and exercises.

The R and RStudio software should be up and running before coming to the workshop. Please visit the Software requirements posted below.

Overview

This workshop introduces the essential concepts of building machine learning models with R. Throughout the workshop you will gain insights in the foundations of machine learning methods, including resampling methods, data preprocessing steps and the tuning of parameters. You will cover a variety of statistical and machine learning methods, ranging from GLMs, over tree-based machine learning methods to neural networks. You will acquire insights in the foundations of these methods, learn how to set-up the model building process, and focus on building a good understanding of the resulting model output and predictions.

Leaving this workshop, you should have a firm grasp of the working principles of a variety of machine learning methods and be able to explore their use in practical settings. Moreover, you should have acquired the fundamental insights to explore some other methods on your own.

Schedule and Course Material

The schedule will gradually be completed over the next days. The detailed schedule is subject to small changes.

Module	Session	Duration	Description	Lecture material	R script	R solutions
Prework		your own pace	check the Prework and Software Requirements instructions
Prework		day before the workshop	download the course material from the GitHub repo
Module 3	Session 1	09.30 - 10.10	Toolbox and tensors	sheets	script
		10.10 - 10.30	MNIST data	sheets	script
		10.40 - 11.10	De-mystifying neural networks	sheets	script
		11.10 - 12.30	Neural networks in {keras}	sheets	script
Module 3	Session 2	9.30 - 10.10	Claim frequency with neural nets	sheets	script
		10.10 - 10.30	Adding a skip connection	sheets	script
		10.30 - 11.10	CNNs	sheets	script
		11.20 - 11.30	Autoencoders	sheets	script
		11.30 - 12.30	Working on case study

Module 3: Neural networks

In two sessions we cover:

• R and Python toolbox
• tensors and operations on tensors
• basics of feed-forward artificial neural networks
• an architecture with layers in {keras}
• loss function, forward pass and backpropagation
• performance metrics
• claim frequency and severity modelling with neural networks
• auto encoders
• convolutional neural networks.

Download lecture sheets in pdf.

A recent working paper on the use of deep learning structures for the analysis of claim frequency and severity data is https://arxiv.org/abs/2310.12671. The notebooks to reproduce the illustrations in the paper are https://github.com/freekholvoet/NNforFreqSevPricing.

Software Requirements

You have two options to join the coding exercises covered during the workshop. Either you join the posit cloud workspace dedicated to the workshop, and then you’ll run R in the cloud, from your browser. Or you use your local installation of R, Python and RStudio.

We kindly ask participants to join the posit cloud as default!

posit cloud

You will join our workspace on posit cloud. This enables a very accessible set-up for working with R in the cloud for the less experienced user!

https://rstudio.cloud/spaces/109060/join?access_code=s3jVNlJn6J8SlZpSXmwvBsCMUC3XicxoFv5kCbCO

Here are the steps you should take (before the workshop):

• visit the above link
• log in by creating an account for posit cloud or by using your Google or GitHub login credentials
• join the space
• at the top of your screen you see ‘Projects’, click ‘Projects’
• with the ‘copy’ button (on the right) you can make your own version of the ‘Module 3 - Neural Networks’ project (created by Roel Henckaerts); in this copy you can work on the exercises, add comments etc.
• you should now be able to visit the project and see the ‘scripts’ and ‘data’ folders on the right
• we already compiled the ‘installation-instructions.R’ script from the scripts folder for you; you do NOT have to run these again; instead open the ‘1_getting_started.R’ instructions and check if these steps work.

We have everything set up for you in the correct way. You only have to login!

Local installation

Alternatively, you can bring a laptop with a recent version of R, Python and RStudio installed. Make sure you can connect your laptop to the internet (or download the course material one day before the start of the workshop). You will need:

• R, see https://cloud.r-project.org/bin/windows/base/
• RStudio, see (https://www.rstudio.com/products/rstudio/download/#download)

Please run the instructions in the below script (Part 1 and Part 2!) to install the required packages as well as Python.

Make sure to download the course material (including scripts and data sets) from the GitHub repo one day before the workshop. Use the green button Code > Download ZIP on the GitHub repo website. As a final step, after downloading, unpack the ZIP folder and make sure you can easily access the material (in particular scripts and data sets).

packages <- c("keras", "tensorflow", "reticulate", "tidyverse", "rstudioapi", "gridExtra", "rsample", "mgcv", "recipes")

# Part 1: install all the needed R packages
new_packages <- packages[!(packages %in% installed.packages()[,"Package"])]
if(length(new_packages)) install.packages(new_packages)

if(sum(!(packages %in% installed.packages()[, "Package"]))) {
  stop(paste('The following required packages are not installed:\n', 
             paste(packages[which(!(packages %in% installed.packages()[, "Package"]))], collapse = ', ')));
} else {
  message("All R packages are installed, part 1/2 completed!")
}

# Part 2: install tensorflow and keras

# Download and install Python from https://www.python.org/, e.g. I have Python 3.11 (64-bit); make sure to install the 64-bit version and a version that is compatible with TensorFlow, see https://www.tensorflow.org/install
# Download Anaconda from: https://www.anaconda.com/products/individual

# Execute the following steps after installing Python
library(keras)
library(tensorflow)
library(reticulate)
path_to_python <- "C:/Users/.../AppData/Local/Programs/Python/Python311/python.exe" # replace this path with the path to your Python location
virtualenv_create("r-reticulate", python = path_to_python)
install_tensorflow(envname = "r-reticulate")
install_keras(envname = "r-reticulate")
use_virtualenv("r-reticulate")

# Alternatively, you can try the instructions posted here
# https://tensorflow.rstudio.com/install/

# Run the following set of instructions as a test
# Warning: the first time you run these instructions you may get some warnings/errors
#          then run the instructions a second time and verify if the ' ... part 2/2 completed!' message appears
if(as.array(tensorflow::tf$abs(-10)) == as.array(keras::k_abs(-10))){
  message('Installation of tensorflow and keras went well, part 2/2 completed!')
} else {
  stop('Something went wrong with installation of tensorflow and/or keras')
}

These instructions can also be downloaded as .R script from the scripts folder. Open this script in RStudio (via File > Open file), select the code and press the Run button top right in the script window.

Instructors

Katrien Antonio is professor in insurance data science at KU Leuven and associate professor at University of Amsterdam. She teaches courses on data science for insurance, life and non-life insurance mathematics and loss models. Research-wise Katrien puts focus on pricing, reserving and fraud analytics, as well as mortality dynamics.

Jonas Crevecoeur co-designed the workshop with us, but will not teach the October - November 2023 edition. Jonas obtained his PhD within the insurance research group at KU Leuven and holds the degrees of MSc in Mathematics, MSc in Insurance Studies and MSc in Financial and Actuarial Engineering (KU Leuven). Before starting the PhD program he worked as an intern with QBE Re (Belgium office) where he studied multiline products and copulas. Jonas was a PhD fellow of the Research Foundation - Flanders (FWO, PhD fellowship fundamental research).

Roel Henckaerts holds the degrees of MSc in Mathematical Engineering, MSc in Insurance Studies and Financial and Actuarial Engineering (KU Leuven) and PhD in Business Economics (KU Leuven). Before starting the PhD program he worked as an intern with AIG (London office) and KBC. Roel is PhD fellow of the Research Foundation - Flanders (FWO, PhD fellowship strategic basic research). After the completion of his PhD, Roel joined Prophecy Labs: an AI/ML startup with experience in building end-to-end data solutions that provide concrete business value.

Happy learning!

---