A data scientist uses logistic regression to build a fraud detection model. While the model accuracy is 99%, 90% of the fraud cases are not detected by the model. What action will definitively help…

Question

Options

• AUsing undersampling to balance the dataset
• BDecreasing the class probability threshold
• CUsing regularization to reduce overfitting
• DUsing oversampling to balance the dataset

• A
  9% (2)
• B
  86% (19)
• C
  5% (1)

Explanation

Decreasing the class probability threshold (B) is the only action that definitively guarantees more fraud cases are detected without retraining the model. Logistic regression outputs a fraud probability for each transaction; the default decision boundary is 0.5, meaning a case is flagged as fraud only if P(fraud) > 0.5. With severe class imbalance (fraud is rare), the model rarely predicts high probabilities for fraud, so lowering the threshold - say to 0.2 - directly forces more cases to be classified as fraud, immediately increasing recall at the cost of more false positives.

Why the distractors fail:

• A (undersampling) and D (oversampling) both require retraining the model and may improve recall, but neither definitively guarantees it - results depend on the data and new model behavior. The question asks for what will definitively help.
• C (regularization) addresses overfitting, which is not the problem here; the issue is a biased decision threshold caused by class imbalance, not model complexity.

Memory tip: Think of the threshold like a "sensitivity dial" on a smoke detector - turning it down (lowering the threshold) makes it trigger on less smoke, catching more real fires but also more false alarms. The dial can be adjusted instantly without rebuilding the detector (retraining the model).

No community discussion yet for this question.