Credit Card Fraud Detection

Kaggle dataset

GOAL: predict whether the transaction is fraud or not.
BEST SCORE: 0.99 Accuracy (best model - DBSCAN)

---

Architecture

EDA

The features:

1. Time - Number of seconds elapsed between this transaction and the first transaction in the dataset
2. V1-V28 - Result of a PCA Dimensionality reduction to protect user identities and sensitive features
3. Amount - Transaction amount
4. Class - 1 for fraudulent transactions, 0 otherwise

Dataset is highly unbalanced, the positive class (frauds) account for 0.172% of all transactions (492 of 284807).

Different models

Isolation Forest

Total metrics:

	Test
Accuracy:	0.98
Predicted anomalies:	0.4266 %

One-Class SVM

Total metrics:

	Test
Accuracy:	0.98
Predicted anomalies:	0.4449 %

Local Outlier Factor

Total metrics:

	Test
Accuracy:	0.98
Predicted anomalies:	0.4252 %

DBSCAN

Total metrics:

	Test
Accuracy:	0.99
Predicted anomalies:	0.2826 %