Normal versus abnormal behaviour

Fraud Detection in Python

Charlotte Werger

Data Scientist

Fraud detection without labels

• Using unsupervised learning to distinguish normal from abnormal behavior
• Abnormal behavior by definition is not always fraudulent
• Challenging because difficult to validate
• But...realistic because very often you don't have reliable labels

What is normal behavior?

• Thoroughly describe your data: plot histograms, check for outliers, investigate correlations and talk to the fraud analyst
• Are there any known historic cases of fraud? What typifies those cases?
• Normal behavior of one type of client may not be normal for another
• Check patterns within subgroups of data: is your data homogeneous?

Customer segmentation: normal behavior within segments

Let's practice!

Fraud Detection in Python