This document discusses using machine learning for fraud detection. It outlines how machine learning can provide a scalable, adaptable solution to identify fraud. The machine learning pipeline involves gathering data from user accounts, selecting a model like CatBoost that handles categorical data well, training and evaluating the model, and deploying the trained model to classify new users as fraudulent or not. The goal is to balance avoiding false positives and false negatives when identifying fraud.
17. Deep Learning
- Very Powerful
- Needs lots (millions) of data
Classic ML:
Random Forest
Gradient Boosting
Choosing a Machine Learning Algorithm
17
MODEL SELECTION
Chosen algorithm: CatBoost
18. - Number of trees
- Depth of trees
- Learning rate
18
HYPERPARAMETERS
Example from Titanic data
Source: Wikipedia
19. CatBoost brings:
- Powerful
- Open-source
- Handles categorical data very well
- Automatic overfitting detection
- (Works on GPU)
19
CATBOOST
21. Trade-off between being too severe/ Being too friendly
Avoiding Overfitting
False positives > Innocents wrongly flagged
False Negatives > Fraudsters not detected
EVALUATING PERFORMANCE
Hyperparameter Tuning
21