Databricks Machine Learning Associate Question 582
Select 3A data scientist is working on a machine learning project where they aim to evaluate the model's performance more reliably. They are debating whether to use cross-validation or a simple train-validation split. Which of the following statements correctly describe the benefits or downsides of using cross-validation over a train-validation split?
- A
Cross-validation provides a more robust estimate of model performance compared to a single train-validation split.
- B
Cross-validation is computationally more expensive than a train-validation split.
- C
Cross-validation is unnecessary for very large datasets where a single train-validation split is sufficient to obtain reliable performance estimates.
- D
Cross-validation ensures that the model is trained on all available data, improving test accuracy.
- E
Cross-validation eliminates the risk of data leakage in the dataset.
Show answer and explanation
Correct answers: A, B, C
Explanation
Cross-validation is a powerful technique for obtaining a more reliable estimate of model performance, as it reduces the variance associated with relying on a single train-validation split. However, it is computationally expensive since the model is trained multiple times. For very large datasets, a single train-validation split may suffice for performance estimation. While cross-validation has its benefits, it does not directly improve test accuracy or prevent data leakage, which must be addressed separately.
- A. Correct.
Cross-validation divides the dataset into multiple folds and trains the model on different subsets, which reduces the variance in performance estimates compared to relying on a single random split.
- B. Correct.
Because cross-validation involves training the model multiple times (once for each fold), it is computationally more expensive than a single train-validation split.
- C. Correct.
For very large datasets, a single train-validation split is often sufficient because the size of the training and validation sets ensures reliable performance estimates. Cross-validation may be overkill in such cases.
- D. Incorrect.
While cross-validation ensures that all data is used for both training and validation, it does not improve test accuracy directly. The goal is to better estimate model performance, not improve the model's accuracy.
- E. Incorrect.
Cross-validation does not inherently eliminate the risk of data leakage. Proper data preprocessing and careful handling of data splits are necessary to avoid leakage.