edward0chen / ieee-cis-fraud-detection Goto Github PK

python 3.7.3

numpy 1.16.2

pandas 0.24.2

sklearn 0.20.3

keras 2.2.4

tensorflow 1.13.1

xgboost 0.82

lightgbm 2.2.3

• Put unzipped data/model data in 01-Input Model_Data
• all the model data can be found in google drive
• Generate a simple solution that is good enough for 3rd place (~0.943642 on private LB)

• 890 features
  cd /02-Feature_Enginnering/890features/
python lgb_single_final.py----Also output the result

• 692 features
  cd /02-Feature Enginnering/692features/1.baseline_features_388/
python 1.feature engineering.py
python 2.feature selection.py
python 3.feature engineering.py
cd /02-Feature Enginnering/692features/2.uid_magic_features_301/
python uid4_eng.py
cd /02-Feature Enginnering/692features/3.combine_features_3/
python combine_features_3.py
  

• 890 features
  cd /03-Single_Model/890features/
python lgb_single_final.py
  -----------------CV 95365 LB 9605

• 692 features
  cd /03-Single_Model/692features/
python Lgb_CV9562_LB9597.py
  -----------------CV 9562 LB 9597
  ----------------- tune the parameters the lgb can reach LB 9614
  python Catboost_CV9582_LB9590.py
  -----------------CV 9582 LB 9590
  python NN_CV9518_LB9556.py
  -----------------CV 9518 LB 9556
  

• cd /04-Model Blend/
python model_blend.py
  
• Finally：lgb_0930_0.65_v1
  
• Public:0.967161 Private:0.943642
  

step 01:
lgb_890features_blend_0.65=0.65*lgb_kfold_9614+0.35*lgb_kfold_9605
------LB 9646-----
step 02:
lgb_0930_0.95_v0=0.95*lgb_890features_blend_0.65+0.05*CV9518_NN_LB9556
------LB 9663-----
step 03:
lgb_0930_0.65_v1 =0.65*lgb_0930_0.95_v0.csv+0.35*pred_692_features_blend
while:
pred_692_features_blend=0.65*lgb_cv9562_692features_9597+0.35*CatBoost_cv9582_692features_9590