Fraud-Prediction

Fraud hosts with substantial amount of fraudulent traffic using the impression logs for selected IP addresses

Requirements

1. python 2.7
2. sklearn 0.19 dev version or with following fix
3. numpy
4. pandas

Run

1. Install the requirements for the projects by running
2. jupyter notebook Fraud_Prediction.ipynb to open the Jupyter notebook

Discussions

1. A One class SVM is used for predicting fraudulent traffic (+1)
2. Using bucketed time_stamps as features increases the test accuracy by >20%, which is a good indicator that fraudulent activities is clustered well in time. This fact can be extended to find the botnet networks.
3. Over-classification with SMOTE was tried to balance a highly imbalanced dataset, but was not useful in producing better results, hence ommitted.