Model Development: Databricks Machine Learning Associate Practice Questions

A retailer wants to forecast next week's unit sales for each store. The target is a real-valued number and model quality will be judged by prediction error in the original units sold. Which modeling family best matches this problem?

1. AA classifier, because the target can be binned into categories after training without changing the task definition
2. BA regressor, because the target is continuous and the model should output numeric sales estimates for each store
3. CA clustering algorithm, because grouping similar stores replaces the need to predict a numeric target value
4. DAn association-rule model, because product co-occurrence patterns are equivalent to weekly sales forecasting targets
5. EA topic model, because it can discover latent themes in text and reuse them as numeric forecasts for sales

A linear baseline underfits both the training and validation sets on a complex nonlinear problem. Which change is most likely to reduce bias and improve fit, assuming enough data is available?

1. AReduce model complexity further by removing interaction terms and increasing regularization to simplify the hypothesis class
2. BUse fewer training examples so the model can memorize the smaller sample and therefore generalize more accurately
3. CIncrease model flexibility, such as adding nonlinear features or using a more expressive algorithm family
4. DReplace cross-validation with one random split so the reported score rises even if the model remains too simple
5. EKeep the same simple model and tune only the random seed, because underfitting is usually caused by unlucky shuffling

A regression target was log-transformed before training to stabilize variance. The team wants the training object itself to handle inverse transformation of predictions automatically during inference. Which tool is the best fit?

1. AA StandardScaler applied to the target column, because it always restores predictions back to the original scale internally
2. BA FeatureHasher, because hashing target values lets the model decode them back to raw numeric units during scoring
3. CA GridSearchCV object, because hyperparameter search automatically reverses target transformations after refitting
4. DA TransformedTargetRegressor, because it wraps a regressor and applies the target transform and inverse transform consistently
5. EA KFold splitter, because each fold stores the target transformation and averages the inverse predictions afterward

An engineer wants to tune hyperparameters and perform cross-validation in one object that will refit the best model afterward. Which tool is the correct choice?

1. AKFold, because it both defines the splits and refits the best estimator chosen from the parameter grid automatically
2. BGridSearchCV, because it combines parameter search with cross-validation and can refit the best setting on the full training data
3. CPipeline, because it alone compares parameter combinations and stores best_params_ after choosing a winning fold
4. Dmetrics.get_scorer, because scorer objects coordinate all folds, select the winner, and train the final estimator
5. Ecross_val_predict, because it tunes parameters and returns the best fitted estimator together with holdout predictions

A regressor was trained on log(y) instead of y. The analyst now wants to report RMSE in dollars because the business users think in original currency units. What should happen before computing that RMSE?

1. ACompute RMSE directly on the logged predictions and logged targets, because the units remain dollars after the log transform
2. BStandardize the logged predictions and targets again so RMSE becomes comparable across datasets with different scales
3. CExponentiate the predictions and the corresponding target values back to the original scale, then compute RMSE in dollars
4. DClip negative logged predictions to zero so the metric stays in the valid range for currency-based interpretation
5. EAverage the fold scores first and exponentiate the final RMSE afterward to convert it into original business units

An engineer has four hyperparameters, each with many plausible values, and only enough budget to test a limited number of configurations. Which search method is usually more efficient than exhaustive grid search in this situation?

1. ARandom search, because it can sample a fixed number of candidates without evaluating every possible combination
2. BGrid search, because it always reaches the best model with fewer evaluations when the space has many unimportant parameters
3. CManual search, because changing values by hand guarantees broader coverage of the parameter space than random sampling
4. DFull enumeration, because evaluating the entire Cartesian product is the only budget-aware approach in high dimensions
5. ELeaveOneOut search, because it removes one parameter value at a time until the best setting remains after elimination

A credit-risk team uses a model to rank applicants for manual review. They care most about whether risky applicants receive higher scores than safe applicants across many possible decision thresholds. Which metric is the most suitable primary choice?

1. AROC AUC, because it measures ranking quality across thresholds rather than depending on one fixed classification cutoff
2. BAccuracy, because it is threshold-free and always captures ranking quality better than AUC for scored classifiers
3. CMean squared error, because risk scores should be treated as continuous regression outputs regardless of the labels
4. DSilhouette score, because ranking applicants is equivalent to maximizing cluster separation in the score space
5. EAdjusted Rand index, because applicant ordering is best measured as agreement between predicted and true partitions

A GridSearchCV job evaluates 4 values of alpha and 3 values of max_iter for a model, using 5-fold cross-validation. Ignoring the final refit on all training data, how many model fits occur during the search phase?

1. A12
2. B15
3. C60
4. D75
5. E240