Objective 3.3 Train, Validate, and Test Data Subsets AIP-210 Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   A data scientist is working on a medical diagnosis dataset where a rare disease occurs in only 2% of patients. The scientist wants to use 5-fold cross-validation to tune a model. Which approach is most appropriate?

   • AOversample the minority class before splitting
   • BUse stratified k-fold cross-validation to preserve the 2% disease rate in each fold
   • CUse a single 80/20 train-test split instead of cross-validation
   • DUse standard k-fold cross-validation without stratification
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A machine learning engineer has a small dataset of 2,000 labeled images and needs to tune hyperparameters for a convolutional neural network. The engineer wants to use the data efficiently while getting a stable estimate of model performance during tuning. Which approach is most appropriate?

   • AUse a single 80/20 train-test split and tune hyperparameters on the test set
   • BUse the entire dataset for training and report the training accuracy as the model's expected performance
   • CUse 5-fold cross-validation on the training portion to tune hyperparameters, then evaluate the final model on a held-out test set
   • DUse a 70/15/15 train-validation-test split and tune on the validation set
   Reveal Solution Discussion Select an answer first
4. 18foundation · easy

   What is a primary advantage of k-fold cross-validation over a single validation split?

   • AIt eliminates the need for a test set
   • BIt uses the entire dataset for training and validation, providing a more reliable performance estimate
   • CIt requires less computational time than a single split
   • DIt always produces higher accuracy than a single split
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   What is the main purpose of temporal splitting for time-series data?

   • ATo use the most recent data for training
   • BTo randomly shuffle the data to break time dependencies
   • CTo respect chronological order and avoid look-ahead bias
   • DTo ensure each subset has the same number of samples
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A financial institution is building a model to predict stock price movements using historical daily data. The data scientist splits the data randomly into 80% training and 20% test sets. The model performs exceptionally well on the test set, but when deployed, its performance drops significantly. What is the most likely cause of this discrepancy?

   • ARandom splitting introduced look-ahead bias, so the test set was not representative of future data
   • BThe model is overfitting the training data
   • CThe test set is too small to provide a reliable estimate
   • DThe model needs more training data to generalize
   Reveal Solution Discussion Select an answer first