Explain fraud types and evaluate a fraud model

You are being interviewed for a Fraud Data Scientist role at PayPal. Respond to the following:

1. Name common fraud types relevant to payments (for example, account takeover, first-party fraud, third-party fraud, merchant fraud). For each one, provide a brief definition and an example.

2. Account Takeover (ATO): Walk through how ATO usually unfolds from start to finish (attacker acquisition → credential takeover → monetization), and describe the signals or features you would expect to help detect it.

3. First-party vs. third-party fraud:

   • Give a clear definition of each.
   • Explain why label quality or ground truth may differ between the two.
   • Describe at least one way mislabeling might bias model training or evaluation.

4. How would you assess the effectiveness of a fraud model in production?

   • Give at least five metrics, covering both ML metrics (such as precision/recall) and business or risk metrics (such as fraud dollars prevented and false-positive cost).
   • Explain the main tradeoffs (for example, precision versus recall) and how decision thresholds should be selected.
   • Make clear what the “positive class” is and how class imbalance affects evaluation.

5. You are asked to build a fraud strategy from the ground up. Outline a practical plan covering data, modeling, decisioning, monitoring, and iteration. Assume you need to deploy something useful within 6–8 weeks.

State any assumptions you require (for example, time zone, label delay, chargeback window).

Overview: This question tests knowledge of payment fraud categories, how account takeover works end to end, labeling and ground-truth issues that separate first-party and third-party fraud, how to choose detection signals and evaluation metrics, and how to design a rapid production strategy; category/domain: Machine Learning, position type: Data Scientist, abstraction level: applied systems and feature-level modeling with operational deployment considerations. It is often asked because payment platforms must balance fraud losses against customer friction and operational cost, so interviewers explore understanding of attack flows, label bias, ML and business metrics (precision/recall, fraud dollars prevented, false-positive cost), threshold tradeoffs, and practical plans for data, modeling, decisioning, monitoring, and iteration under tight timelines and label/chargeback delays.