FraudDetectionModel

In this notebook I use a publickly available dataset 'FraudDetection.xlsx'to train fraud detection models.

• See Notebook fraud_detection_imbalance_classes in this repo for Weighted Logistic Regression.

Two Fraud Detection models trained:

1. Logistic Regression
2. AdaBoost Classifier

The accuracy of both the Logistic Regression and AdaBoost Classifier models using the test data are greater than 85%!! This suggests that the model is really good but can further be improved by tuning some hyper-parameters.

The CAP curve analysis shows that that if 50% of the data is analysed without the model only 50% of the fraudulent cases would be discovered. But using the model, we would be able to identify more than 90% of the fraudulent cases in the same amount of data!!

Using the model:

• the amount of backlog (incoming and unfinished work) would be reduced significantly,
• the amount of staff needed to do the job would be reduced drastically,
• the processing time would be shortened significantly and
• more cases of fraudulent transactions would be tracked down in a given amount of data processed - more than 40% increase in effeciency!