Question: A team is training a churn predictor with L2 weight decay and dropout, and the model appears to be underfitting. Explain why reducing or removing these regularization methods can improve training fit, and why that same change often makes validation performance less stable. What should the team check before changing the regularization setting?

Sample answer: Regularization limits the set of patterns a model can learn, which helps prevent it from fitting noise but can also keep it from fitting the training data well enough. If the model is underfitting, relaxing those constraints gives it more freedom to match the training examples more closely, so training error can fall and the bias part of the error can shrink. The downside is that a more flexible model can start capturing accidental patterns in the training set, so its predictions become more sensitive to the exact data it saw and validation error may rise. Because of this trade-off, the team should first confirm that underfitting is the main problem rather than overfitting before reducing regularization.

Key points:

• Regularization reduces model freedom
• Less regularization can improve training fit and reduce bias
• Less regularization can increase variance and validation sensitivity
• The choice depends on whether underfitting or overfitting is the bigger issue

Rubric: Full credit requires explaining how regularization constrains model capacity, why easing it can reduce bias, why it can increase variance, and why the adjustment should be made only after checking which error source dominates.