Predicted fraudulent events for Eventbrite using event data with an accuracy of 98.8% and a recall of 91.1%. Created a web app that classifies fraudulent events on streaming data.
Original Data: 44 features, Cleaned Data: 8 features
- Dropped collinear features
- Dropped features with weak correlation
- Dummied categorical features
- Aggregated costs in ticket_types for total_cost
- Extracted # previous payments from previous payouts feature (most frauds had 0 previous payments while others had 10+)
- Imputed null values with mean (selected features had low # of null values)
- Implemented this in
munging.pyto createclean.csv
delivery_method, sale_duration, num_order, num previous_payouts, user_type, fb_published, channels, total_cost
Accuracy: 0.9880, Recall: 0.9112
| Predicted Not Fraud | Predicted Fraud | |
|---|---|---|
| True Not Fraud | 3265 | 16 |
| True Fraud | 27 | 277 |
Accuracy: 0.9704, Recall: 0.8500
| Predicted Not Fraud | Predicted Fraud | |
|---|---|---|
| True Not Fraud | 3207 | 58 |
| True Fraud | 48 | 272 |
Cleaned data with munging.py and saved it as a csv file. Created classifier in model.py and pickled it for the web app. Modified munging.py to work with one row of data at a time (for streaming data) and saved it as munging_stream.py. Created a web app with Flask in app.py.