Databricks Machine Learning Associate Question 598
Select 2You are working on a binary classification problem to predict whether a customer will make a purchase (Yes/No). After training your model, you evaluate its performance using various metrics. The following results are obtained:
- F1 Score: 0.85
- ROC/AUC: 0.92
- Log Loss: 0.35
Which metric(s) would you prioritize if the cost of false positives is significantly higher than false negatives?
- A
F1 Score
- B
ROC/AUC
- C
Log Loss
- D
Precision
- E
Recall
Show answer and explanation
Correct answers: C, D
Explanation
When false positives are costlier than false negatives, metrics that specifically address the cost associated with false positives, such as Precision and Log Loss, should be prioritized. Precision directly measures the tradeoff related to false positives, while Log Loss accounts for the confidence of predictions, making it suitable for cost-sensitive decisions. Metrics like F1 Score and ROC/AUC provide valuable insights but are not tailored to address this specific cost-sensitive scenario.
- A. Incorrect.
The F1 Score is a harmonic mean of precision and recall and is a good metric for imbalanced datasets. However, it does not specifically address the scenario where false positives are costlier than false negatives.
- B. Incorrect.
ROC/AUC measures the ability of the model to distinguish between classes across different thresholds, but it does not focus specifically on the cost of false positives or negatives.
- C. Correct.
Log Loss penalizes incorrect predictions by assigning higher penalties for incorrect confidence levels. This metric is particularly useful when you want to account for the probability predictions, making it relevant when cost-sensitive decisions are required.
- D. Correct.
Precision focuses on reducing false positives by calculating the proportion of true positive predictions among all positive predictions, making it the appropriate metric when false positives are costlier.
- E. Incorrect.
Recall measures the proportion of true positive predictions among all actual positives, which is more relevant when false negatives are the priority rather than false positives.