Machine Learning & Predictive Analytics Flashcards

What is the purpose of a holdout validation set during model development?

Answer: To tune hyperparameters without touching the test set

A validation set is used to tune hyperparameters so the test set remains an unbiased final estimate.

In gradient descent, what role does the learning rate play?

Answer: It controls the step size of each parameter update

The learning rate determines how large a step the algorithm takes toward minimizing the loss each iteration.

Which metric would be most misleading for a dataset that is 95% negative class?

Answer: Accuracy

Accuracy can look high (95%) by always predicting the majority class, hiding poor minority-class performance.

What is the F1 score a measure of?

Answer: The harmonic mean of precision and recall

The F1 score is the harmonic mean of precision and recall, balancing the two.

Why might one use a time-based split instead of a random split for a forecasting model?

Answer: To avoid training on future data and mimic real deployment

Time-based splits prevent the model from learning on future data, reflecting how it will actually be used.

What does feature importance from a tree-based model tell you?

Answer: How much each feature contributes to reducing impurity

Tree-based feature importance reflects how much each feature reduces node impurity across the model, not causation.

A stakeholder needs to understand exactly why a model made each prediction. Which model type is most interpretable?

Answer: Logistic regression

Logistic regression offers transparent, coefficient-based explanations that are easy to interpret.