Credit Card Fraud Detection

1. Came across this mocked-up dataset of customer transactions at [Capital One Recruitment Challenge](https://github.com/CapitalOneRecruiting/DS).
2. The unbalanced dataset is comprised of artificial customer transactions with a few outlier cases where fraud was detected. There's only ~1.6% fraudulent cases.
3. Our primary goal is to successfully predict whether a transaction is Fraudulent or not, and avoid Type-II errors as much as possible as in most sensitive classification problems: we'll try not to point accusatory-fingers at genuine-transactions 🙂.
4. The secondary goal is to identify interesting anomalies in the transactions like multi-swipes, reversal of suspicious transactions, etc. by performing exploratory-data-analysis.
5. Most numerical-fields seem to follow Power-law distributions rather than Gaussian distributions.
6. We'll engineer some time-dependent categorical features by parsing the datetime fields, exclude the fields which have just one categorical value (makes no sense keeping these around 🙁), and also create a new feature to indicate if credit-card-CVV is wrongly entered.
7. Baseline classifiers chosen are Logistic Regression, Decision Tree, and Isolation Forest.
8. Performance is kinda poor on these Baseline classifiers: precision, and recall vary greatly across the models. Logistic Regression just predicts the majority class (ie- transaction is OK).
9. Final Random-Forest achieves a Recall-score of approximately 99.99% indicating that False-Negatives are absolutely minimal.