Credit Card Fraud Detection

Fraud detection on a dataset that is 0.172% positive — where accuracy is a useless metric and the whole problem is resampling strategy and the recall/precision trade-off.

Conducted fraud detection on the Kaggle dataset (284,807 transactions, only 492 fraud → 0.172%).
Applied SMOTE and ADASYN to handle severe class imbalance.
Implemented and compared Logistic Regression, Decision Tree, Random Forest, and XGBoost.
Achieved ROC-AUC 0.9817 and Recall 0.735 with XGBoost on imbalanced data.
Improved Recall to 0.816 with ADASYN while maintaining high ROC-AUC (0.9744), outperforming other models.
Contributed to data preprocessing, resampling, model development, and evaluation, delivering a robust prototype for financial fraud detection.

Explain this for

ROC-AUC: 0.9817; XGBoost on imbalanced data
Recall: 0.816; with ADASYN
Positive rate: 0.172%; 492 frauds in 284

The problem

The Kaggle credit-card dataset contains 284,807 transactions of which 492 are fraudulent — 0.172%. A classifier that predicts "legitimate" unconditionally scores 99.83% accuracy and catches nothing. This is the canonical case where the headline metric actively misleads, and every meaningful decision in the project follows from refusing to use it.

Resampling

I compared SMOTE and ADASYN. Both synthesise minority-class examples by interpolating between neighbours, but they choose where to synthesise differently: SMOTE spreads uniformly across the minority class, ADASYN concentrates on the examples that are hardest to classify — the ones sitting near the decision boundary, which is where fraud actually lives.

Models

Four classifiers, on both the raw and resampled distributions: Logistic Regression, Decision Tree, Random Forest, XGBoost.

Setting	ROC-AUC	Recall
XGBoost, imbalanced	0.9817	0.735
XGBoost + ADASYN	0.9744	0.816

Reading the trade-off

This is the entire finding, and it is a trade, not a win. ADASYN lifts recall from 0.735 to 0.816 — roughly 40 additional frauds caught out of every 492 — at the cost of 0.7 points of ROC-AUC.

Which side of that trade you want is a business question, not a modelling one. A missed fraud costs the issuer the transaction value plus investigation overhead; a false positive costs a declined card and an annoyed customer. The ratio between those decides the operating point, and the honest output of this project is the curve rather than a single chosen threshold.

What I took from it

Choosing the metric is the modelling decision. Once ROC-AUC and recall replaced accuracy, every subsequent choice — which resampler, which model, which threshold — became answerable with evidence. Before that, the numbers were all excellent and all meaningless.