Can we predict if people will choose a certain option out of a set?

This repository contains data files and scripts to run machine learning algorithms using the function automl from h2o.ai in R.

Documentation from h2o ai can be found at: https://docs.h2o.ai/h2o/latest-stable/h2o-docs/automl.html

See https://hackmd.io/@dlholf/manuscript#About-the-dataset for explanation about the dataset.

Data is contained in the folder 'data'.

There are 5 script files in the folder 'scripts'.

1. data_clean.R cleans and combines the raw data file into a form ready for analysis.
2. add_strategies.R codes participants' choices according to whether it is:

• The best equally-weight average by %RI
• The best traffic light bands (on average)
• Minimising one of the following nutrients' %RI: sugar, salt, energy, fat
• Minimising one of the following nutrients' TFL band: sugar, salt, fat

3. run_models.R runs the h2o.automl models for 9 different criteria (5 based on %RI, 4 based on TFL bands) and extracts metrics (SHAP values etc.) from each model. It also plots and saves SHAP and PD plots.

4. surrogate_plots.R does the same as 3, but with 'Yes'/'No' coding of the binary choices, such that we can plot surrogate trees using these responses instead.

5. metrics.R analyses the extracted metrics from 3.

The folder 'plots' contain the SHAP and partial dependency plots produced in the scripts.

The folder 'metrics' contain the extracted model fit metrics and SHAP values produced in the run_models.R script.