Datathon 2021 Repository
- notebooks
- raw_data <-- do NOT check in; keep in your local folders only
|- DataDictionary*.csv
|- Training*.csv
| Who | Submission | Score | Notes |
|---|---|---|---|
| Jeny | 1 | .84172 | Catboost |
| Jeny | 2 | .49744 | KNNImputer, Catboost |
| Jeny | 3 and 4 | .50456 | |
| Jeny | 5 | .49744 | Remove lower features from 2 |
| Jeny | 6 | .85176 | AutoViML |
| Jeny | 7 | .83502 | Hour Data, AutoViML |
| Jeny | 8 | .85630 | Day Data, AutoViML |
| Jamie | 1 | .616 | RandomForest no FE |
| Jamie | 2 | .841 | Tabnet v2 |
Submission 1 In this submission I dropped all columns with over 40% missing values and ran through Catboost. Score on Kaggle: 0.84172
Submission 2 In this submission, I deleted 72 features determined not useful after looking at charts from my SweetViz notebook and/or because of excessive missing data. I then did feature engineering to include but reduce the categories for hospital_id and icu_id, counting their frequencies of encounters and binning hospitals and icus into 1-3 and 1-5, respectively smallest to largest. These new features are named hospital_size and icu_size Next I used KNNImputer to impute the missing data. Finally the model was again trained with Catboost. Score: 0.49744
Submission 3/4 fail in code and setting up Score: 0.50456
Submission 5 Removed lower 89ish of feature importances from Sub 2 Score: 0.49744
Submission 6 Trying out featurewiz and autoviml Score: 0.85176