lakshanweerasinghe/ml_project

Preprocessing techniques

1. Imputed following categorical features with most frequent value quantity,management_group,source_class, basin, payment, payment_type, permit, quantity_group, water_quality, quality_group, region, extraction_type_group, extraction_type, source, source_type, waterpoint_type, waterpoint_type_group, scheme_management, subvillage, ward, wpt_name

2. Imputed following categorical features with constant lga, installer, funder,extraction_type_class, management.

3. Imputed all numerical features with mean.

4. Scale all numerical features using StandardScaler

Feature Engineering techniques

1. Created new features year_recorder, yearly_week_recorder, month_recorder using date_recorded and construction_year.

2. Created a new feature age by subtracting date_recorded and construction_year.

3. Created two new features distance and angle using latitude and longitude.

4. Created two new features distance_pca0 and distance_pca1 by applying Principal Component Analysis to latitude and longitude.

5. Applied OneHot Encoding for following features quantity,management_group,source_class

6. Applied Target Encoding for following features lga, installer, funder,extraction_type_class, management

7. Applied Label Encoding for following features basin, payment, payment_type, permit, quantity_group, water_quality, quality_group, region, extraction_type_group, extraction_type, source, source_type, waterpoint_type, waterpoint_type_group, scheme_management, subvillage, ward, wpt_name

8. Mutual Information scores used to select most appropriate features to train the model as feature selection technique.

Model Selection

1. Initially used XGBClassifier, RandomForestClassifier and CatBoostClassifier. Finally selected RandomForestClassifier as the model to do experiments as it give best results in the selection process.

Proof of submission

Final Rank

1643

Final Score

0.8169