Rendiere/zindi-sa-covid-19-vulnerability-hackathon

My solution the Zindi South African COVID-19 Vulnerability Map Hackathon. # South African COVID-19 Vulnerability Map My solution to the hackathon on Zindi for predicting a vulnerability map in South Africa ## Method Overview I tried a few different things, but the final (best scoring) approach looked something like this: 1. Remove outliers and highly influential points using OLS. 2. Basic feature engineering - see `3. Create Data.ipynb` 3. Feature Selection - Train LightGBM model on full dataset and select top N features. 4. Train some base models and stack them using simple averaging. Most of the modelling was done in `4.3 Modelling.ipynb` so if you want to go follow along then focus there. The stacking process was inspired by this kaggle kernel. It was a first time using a stacking approach for me, don't really like the ideology behind it, but thought I'd give it a try. ## Result I teamed up with the only other South African who was featuring on the top 50 in public leaderboard - Tshepo Maogi, and together we placed **33rd out of 179** with a RMSE of 4.0916 (#1 was ~3.51...).