For our first ML project, our group worked on a Zindi challenge for Credit Fraud. The stakeholders requirement was to maximise the F1-Score metrics.
- dealing with highly imbalanced data (only 0.2% of all cases were fraudulent)
- we used SMOTE algorithm for re-scaling
- create additional insights through feature engineering
- especially, handling date-time object
- work out behavioral patterns from the given data
- dummy classifier to guess minority class (maximises F1-Score) as a base line model
- decision tree
- random forest
- AdaBoost
- stacking (decision tree, random forest, AdaBoost, meta: logistic regression)
pyenv local 3.9.4
python -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt