How would you build and evaluate a classifier?

Scenario: Fraud Detection System

You are tasked with building a binary classifier to detect fraudulent credit card transactions for a global banking platform.

Requirements

1. Data Context: The dataset contains 10 million transactions per day, with a fraud rate of approximately 0.05%. The data is heavily imbalanced.
2. Business Goal: You must minimize financial loss due to fraud while keeping the 'False Positive' rate low to avoid annoying legitimate customers.

Task

Design the end-to-end approach to solving this problem by addressing the following:

• Evaluation Metrics: Given the class imbalance, explain why Accuracy is an inappropriate metric and justify the selection of your primary optimization metric. How would you choose the classification threshold?
• Model Development: Describe your strategy for handling class imbalance (e.g., resampling, cost-sensitive learning) and feature engineering steps for transaction data.
• Deployment & Monitoring: How would you monitor the model's performance in production? Specifically, how do you handle 'model drift' when transaction patterns change?
• Interpretability: If a customer disputes a transaction marked as fraud, how would you use methods like SHAP to explain the model's decision, and what are the limitations of these explanations in a high-stakes financial context?

Expected Output

Provide a structured explanation covering your methodology for the points above, ensuring you highlight the trade-offs between precision and recall relative to the business goal.