Thanks @Amy_Bray and whole team @Zindi for everything you do. Congratulations also to all participants. The following is solo team Sinking Sand approach. Team name was adopted in anticipation of a major leaderboard shakeup that to some extent happened. Models not grounded on solid foundations would shake. I thought mine would be shaken badly and sink down the leaderboard-atleast now I see how lucky I was! Thanks to Zindi not including Cholera predictions in both public and private leaderboard evaluations.

INTRODUCTION

• All datasets are used
• No data reduction is done
• Only one model is used-Lightgbm
• No Post Processing is done to predictions

DATA CLEANING

• Duplicates: keep only the row with the maximum 'Total' value
• Drop nulls in water_sources-around 60 nulls
• Keep only the first four segments of facility ID to harmonize the column in both train and test
• Drop Lat and longitude features-did not have any predictive importance

MERGE

• Get anomalies & percent changes of all climate features in the extra datasets after grouping by 'lat_lon' before merging with train and test. These would later be powerful features for our models when we merge with train and test.
• Ckdetree was used to merge the data with the extra datasets
• Get 3 most nearest points to our location & also calculate their euclidian distances from our location.
• Integrate the climate features plus the anomaly and percent features of these three nearest points to our location

FEATURE ENGINEERING

• Monthly cyclical features-to capture the cyclical nature of months
• Used distances from the 3 nearest points from the location to create features such as Closest distance to water, toilet, and waste facilities of a certain location
• Group target encodings e.g. train.groupby(["Location", "Disease"])["Total"].median(). The problematic location (ID_0358ea0e-af2d-450e-8124-4144ca7860fa) in test but not in train was left as is (with NaNs)
• Geographical clustering using kmeans.

MODELLING

• Many models were tried initially, Lightgbm Regressor worked best comparatively
• Simple Train Test Splitting worked best comparitively
• Log transforming target was better than having raw targets
• A different fitting is done for different diseases
• On cross-validation:Diarrhoea had the highest MAE while dysentry had the least
• On test: Public Lb-around 5.6, Private LB-around 6.6
• No post processing was done

FINAL THOUGHTS

• Location & Disease grouped target encoding was the best feature in all disease scenarios
• I still feel data reduction if explored could have improved the results further
• Tuning lgbm parameters could improve the results further. Was not able to use provided credits due to errors-did not send an email for help though, my mistake!