ari-schaffel/creditcardfrauddetection

Training models on skewed data

CreditCardFraudDetection

Using the anonymous data of credit card transactions available on Kaggle.

The data has about 290000 rows and 31 features. Less than 0.2% of these transactions are fradulent. The goal is to learn a model to predict these fradulent transactions. This is finding the needle in a haystack problem.

In the CreditCardFraudDetection.ipynb notebook, we train logistic regression and SVM through resampling (specifically undersampling) method. The parameters of the models are chosen through KFold cross validation to give a high recall accuracy (we want to minimize false negatives).

For unseen data, logistic regression and SVM both give about 93% recall accuracy, with SVM doing slightly better.

ROC analysis on the test sample gives area under curve = 0.97. The effect of changing threshold on the confusion matrix and the precision recall curve is also investigated.

In CCFraud-AnomalyDetection.ipynb notebook, we use Anomaly detection algorithm by modeling joint probability distribution of features. The probability distribution is modeled on training data which contains only normal transaction examples. The threshold is decided by looking at the cross-validation set which contains both normal and fraud transaction. The resulting threshold choice is then evaluated on the test set.