Question 15
Domain 3: Model DevelopmentOnly 5% of the training examples are fraud cases. Which technique directly addresses class imbalance during model training?
Correct answer: A
Explanation
Assigning a higher class weight or cost to the minority class makes fraud errors count more during training, so the model pays greater attention to the rare fraud examples. This directly addresses class imbalance by changing the loss function to penalize misclassifying the minority class more heavily.
Why each option is right or wrong
A. Assign a higher class weight or cost to the minority class
With only 5% fraud, the training set is highly imbalanced, so the loss function should be modified to penalize mistakes on the minority class more heavily. In practice, weighted cross-entropy or class-cost weighting assigns a larger multiplier to fraud labels during optimization, which directly changes the gradient contribution of those 5% examples and is the standard way to address imbalance at training time.
B. Drop all majority-class examples except one
Keeping only one majority example destroys useful signal and severely undertrains the model.
C. Replace all numeric features with strings
Changing numeric features to strings alters data types, not class balance or loss weighting.
D. Remove the label column before training
Removing labels prevents supervised training because the model no longer knows the target outcome.