Business problem: Given anonymized transaction data with 190
features for 500000 American Express customers, the objective is
to identify which customer is likely to default in the next 180 days
Solution: Ensembled a LightGBM 'dart' booster model with a 5-layer deep CNN. Both models involved significant feature engineering with the LightGBM
model optimized for minimizing logloss and the CNN(activation: Mish) model optimized for reducing focal loss. Used weight of evidence encoding to generate new features for the LightGBM model and calculated payment/balance statement related features for the Keras CNN.
Results: Ranked in the top 7.5% amongst 4875 data scientists and netted a bronze medal. Advanced to 'Competitions Expert' tier on Kaggle.
ol0fmeistr / american-express-credit-default Goto Github PK
View Code? Open in Web Editor NEWIdentify American Express customers most likely to default in the next 3 months based on 190 anonymized transaction data features for over 500000 American Express customers.