Navigating Bias-Variance and Imbalanced Classes in Fraud Detection

Requirements Clarification & Assessment

1. Understanding the Business Context:

   • Objective: Develop a model to predict fraudulent transactions using historical transaction data.
   • Output: Binary classification (0 = legitimate, 1 = fraudulent).
   • Goal: Minimize financial losses by accurately identifying fraudulent transactions.

2. Data Characteristics:

   • Time Span: A decade's worth of transaction data.
   • Class Imbalance: Fraudulent transactions constitute 0.01% of the dataset.
   • Features:
     • Transaction amount
     • Merchant category
     • Merchant zip code
     • Billing address zip code
     • Average transaction amount for the past six months

3. Model Requirements:

   • Real-time Prediction: Ability to flag transactions in real-time for review.
   • Evaluation Metrics: Prioritize recall and F1-score due to the cost of missing fraud.
   • Scalability and Robustness: Handle large volumes of transactions efficiently.

4. Additional Considerations:

   • Segmentation: Whether different models are needed for different card types or user segments.
   • Periodicity and Seasonality: Consider changes in spending habits over time.