This repo includes my Bronze Medal solution (Top 10% finish) to the Kaggle competition hosted by IEEE-CIS to detect fraud transactions.
Stacking & Blending Strategy :
A. Easiest - Weighted blending of top scoring public kernels - https://www.kaggle.com/priteshshrivastava/ieee-cis-blend
B. Linear stacking - https://www.kaggle.com/aharless/simple-linear-stacking-lb-9730
C. Stacking with a meta model :
- Create a pipeline of kernels starting from data loading & reducing memory usage
- Create a hold out validation set
- should we split this by time?
- https://www.kaggle.com/kyakovlev/ieee-cv-options
- Create single models reading the split data and predicting on val & test sets
- A. Catboost - https://www.kaggle.com/priteshshrivastava/ieee-pipeline-2-a-model-a-catboost-feat-sel
- Errors due to missing values / data type
- B. Random Forest - https://www.kaggle.com/priteshshrivastava/ieee-pipeline-2-b-model-b-random-forest
- C. XGBoost - https://www.kaggle.com/priteshshrivastava/ieee-pipeline-2-c-model-c-xgboost
- A. Catboost - https://www.kaggle.com/priteshshrivastava/ieee-pipeline-2-a-model-a-catboost-feat-sel
- Create a meta model that trains on val preds and predicts on test set
- [Optional] Blend this output with other kernels (?)
- useful kernels : https://www.kaggle.com/yw6916/lgb-xgb-ensemble-stacking-based-on-fea-eng