A model must flag fraudulent card transactions, which are a tiny fraction of all activity. Which evaluation approach suits this problem?
Show answer and explanation
Answer: B
Fraud detection is a classification task with a heavily skewed class balance, so the metrics must focus on the minority class. Precision reports what share of flagged transactions were genuinely fraudulent, and recall reports what share of all fraud the model caught. Looking at both exposes the trade-off a single threshold creates. MAE, MSE and R-squared measure the size of numeric errors, which only makes sense when the target is a continuous quantity.
Why the other options are wrong
- A. Mean absolute error averages the magnitude of numeric residuals, so it applies to regression rather than a fraud/not-fraud label.
- C. R-squared describes how much variance a regression explains; it is undefined for categorical outcomes.
- D. Mean squared error penalises numeric prediction error and cannot express missed fraud or false alarms.
Exam tip. Skewed classification: quote precision and recall (or their harmonic mean, F1), never accuracy alone.