Account Takeover Prediction System

Problem Statement

Design a machine learning system to predict account takeover (ATO) risk for a payments API platform. Account takeover occurs when bad actors gain unauthorized access to legitimate user accounts through stolen credentials or identity fraud.

Key Design Considerations

This problem intentionally leaves the use case open-ended. Before diving into the solution, you should clarify the specific scenario and requirements with the interviewer.

Interview Structure

During the interview, expect deep dives into these areas:

Model Design (~20 minutes)

• What input features would you use? Consider both numerical and categorical features:
  • Account behavior data: login time, geographic location, device type
  • Transaction records: amount, frequency
  • IP address information
• How would you preprocess these features? (Handling missing values, feature standardization, categorical encoding)
• Which models are appropriate for this task? What evaluation metrics would you use and why?
• What risks and challenges might arise during model development? (Data imbalance, feature redundancy, overfitting)

System Implementation (~15 minutes)

• What are the core components of your system architecture?
  • Data collection module
  • Model prediction service (real-time or batch prediction)
  • Data storage and monitoring
• How do you ensure computational efficiency and acceptable response times at scale?
• How do you handle data latency issues?
• What mechanisms would you use for online model updates?
• How do you monitor model performance in production? (Prediction accuracy, system efficiency)

Business Impact (~15 minutes)

• How do you measure the model's real-world impact?
  • Reducing the probability of users being hacked
  • Improving account security and increasing user trust
• What risks come from false positives degrading legitimate user experience?
• How would you propose improvements to maximize business value?

Train/Test Split (Common Follow-up)

• How do you handle class imbalance in your training data?
• What sampling strategies would you use? (Oversampling, SMOTE, stratified sampling)