feat: Add model training notebook (PH)

[tm-jc-nacpil]

Overview

Addresses #72

This PR add the end-to-end initial model training notebook for the Philippines. The notebook processes the features given the PH DHS clusters, prepares the data, performs (simple) cross-validation evaluation, and finally trains and evaluates a random forest model to predict the wealth index

Input

• Input DHS cluster data for the Philippines

Output

• Cached OSM, Ookla, VIIRS datasets
• trained model (not yet saved)
• cross validation metrics (5-fold cross-validation, repeated): mean R^2 of 0.58 (0.04 std)
• final training metrics and feature importance graphs

Notes

• The initial downloading of the datasets takes very long since it's downloading many different datasets to disk. Expect to wait around 2-4 hours while the datasets are being downloaded.
• The final R^2 achieved is much lower than the cross-validated R^2, likely due to chance. Changing the train_test_seed sometimes results in R^2 = 0.58 and higher instead
• I borrowed some of the evaluation functions. We can collect these into a model_utils module in the future.
• Spatial cross-validation is yet to be implemented

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[alronlam]

In summary, I think what's important is we use SHAP for feature importance. Random forest feature importance can be quite unstable. Good to do it soon so that's what we report in our sprint review.

Scroll down to the SHAP part to see simple 1/2-liners on how to produce a SHAP Chart. https://mljar.com/blog/feature-importance-in-random-forest/

Probably do this for later cause it takes more time, but we can also consider creating a modified version of the bar chart where it's colored according to whether (in general) the feature makes the wealth estimate go higher/lower. Reference code for generating simplified SHAP charts from our air quality repo:

• https://github.com/thinkingmachines/unicef-ai4d-air-quality/blob/main/scripts/train.py#L98-L117
• https://github.com/thinkingmachines/unicef-ai4d-air-quality/blob/main/src/modelling/eval_utils.py#L81

The other comments we can probably improve later on :D