- Developed several machine learning models and neural networks using Sklearn and TensorFlow to predict whether a credit card transaction is fraudulent or not with 0.86 recall, 0.86 precision and 0.93 balanced accuracy scores.
- Used hyperopt library for hyperparameter tuning via TPE for LGBoost, XGBoost and CatBoost models which resulted in an improvement of 0.15 in balanced accuracy score.
- Used methods such as Stratified K-Fold and SMOTE-Random Under Sampling to deal with exceedingly imbalanced data.
The notebook can be viewed here.