I85PREP
⌘ K
H2O.ai Certified Machine Learning study materials
H2O.ai Certified Machine Learning study resources

H2O.ai Certified Machine Learning practice questions

One question at a time. Build your H2O.ai Certified Machine Learning understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Machine Learning Fundamentals and Model Evaluation

A model scores 0.98 on its training set but only 0.62 on a held-out test set. Which conclusion best describes this pattern?

  1. The model requires more features to raise training accuracy
  2. The model is underfitting because training error is nonzero
  3. The test set is too large and dilutes the score
  4. The model is overfitting and has poor generalization Correct answer.

Explanation

A large gap between high training performance and much lower held-out performance is the classic signature of overfitting: the model has memorized training-specific noise rather than learning transferable structure. Generalization error is what matters for deployment, so the fix involves regularization, more data, or simpler models.

Question 2Machine Learning Fundamentals and Model Evaluation

A fraud detection dataset has 0.5% positive cases. A model that predicts 'not fraud' for every record reports 99.5% accuracy. Why is accuracy misleading here?

  1. Accuracy rewards the trivial majority-class prediction and hides zero recall on fraud Correct answer.
  2. Accuracy is undefined for binary classification problems
  3. Accuracy can only be computed on the training set
  4. Accuracy ignores the majority class entirely

Explanation

With severe class imbalance, a constant majority-class prediction attains high accuracy while detecting none of the minority cases that matter. Metrics such as recall, precision, F1, or area under the precision-recall curve reveal this failure. Accuracy is defined for binary problems but is simply uninformative under imbalance.

Question 3Machine Learning Fundamentals and Model Evaluation

Which pair of quantities does a confusion matrix report for a binary classifier at a fixed decision threshold?

  1. Training loss and validation loss
  2. Feature importance and coefficient values
  3. Learning rate and batch size
  4. True/false positives and true/false negatives Correct answer.

Explanation

A confusion matrix tabulates predicted versus actual labels, yielding counts of true positives, true negatives, false positives, and false negatives. Precision, recall, and related metrics are derived from these four cells. Learning rate, batch size, and losses are training hyperparameters or optimization signals, not classification outcomes.

Question 4Machine Learning Fundamentals and Model Evaluation

Which statement correctly defines recall for a binary classifier?

  1. The fraction of predicted positives that are actually positive
  2. The fraction of all records classified correctly
  3. The fraction of actual positives that are correctly predicted positive Correct answer.
  4. The fraction of actual negatives that are correctly predicted negative

Explanation

Recall, also called sensitivity or true positive rate, is TP divided by (TP + FN): of all truly positive cases, how many were found. Precision instead divides by all predicted positives, and overall accuracy divides by all records.

Question 5Machine Learning Fundamentals and Model Evaluation

A regression model is evaluated with mean absolute error (MAE). What does an MAE of 4.2 indicate?

  1. Predictions are 4.2% accurate on average
  2. The model explains 4.2% of the variance
  3. Predictions are on average 4.2 units away from the actual values Correct answer.
  4. The largest single error is 4.2 units

Explanation

MAE averages the absolute differences between predictions and actual targets, so it is expressed in the units of the target variable. It is not a percentage, not a variance measure, and not a bound on the maximum error.

Question 6Machine Learning Fundamentals and Model Evaluation

Why is a separate test set kept untouched until final evaluation?

  1. To speed up gradient descent during training
  2. To replace the need for cross-validation entirely
  3. To provide an unbiased estimate of generalization after model selection Correct answer.
  4. To increase the number of features available to the model

Explanation

If the test set influences model choice, its score becomes optimistically biased. Holding it out until the end preserves an unbiased estimate of performance on unseen data, while validation or cross-validation data guides tuning.

Question 7Machine Learning Fundamentals and Model Evaluation

Which metric is most appropriate when the cost of a false negative greatly exceeds the cost of a false positive, as in screening for a serious disease?

  1. Negative predictive value
  2. Mean squared error
  3. Recall (sensitivity) Correct answer.
  4. Specificity

Explanation

When missing a true positive is the costly error, the metric should penalize false negatives, which recall does by measuring the fraction of actual positives detected. Specificity and negative predictive value emphasize correct rejection of negatives, the opposite priority.

Question 8Machine Learning Fundamentals and Model Evaluation

A team tunes hyperparameters on a validation set and reports the best validation score as the final performance estimate. What is the main weakness of this practice?

  1. Validation scores are always lower than training scores
  2. Cross-validation cannot be combined with hyperparameter search
  3. The estimate is optimistically biased because the validation set influenced model selection Correct answer.
  4. Hyperparameters cannot be tuned without a test set

Explanation

Selecting the configuration that maximizes validation performance makes that score a maximum over many candidates, so it overstates expected performance on new data. An untouched test set or nested cross-validation provides a cleaner estimate.

Question 9Machine Learning Fundamentals and Model Evaluation

Which scenario is the clearest sign that a dataset should be split with a time-based rather than random split?

  1. The goal is to forecast future behavior from historical records Correct answer.
  2. The dataset contains fewer than 1,000 rows
  3. The target variable is continuous
  4. Records are independent and identically distributed

Explanation

When the deployment task is forecasting forward in time, training on future data and testing on the past leaks information and misrepresents the task. A chronological split ensures the model is evaluated only on data that would genuinely be unseen at prediction time.

Question 10Machine Learning Fundamentals and Model Evaluation

A preprocessing step standardizes features using the mean and standard deviation of the full dataset before cross-validation. What is the problem?

  1. It makes the model unable to converge
  2. It converts the problem into a regression task
  3. It leaks validation-fold information into training, inflating scores Correct answer.
  4. Standardization is never appropriate for numeric features

Explanation

Computing scaling statistics on all data lets information from the held-out fold influence training, a form of data leakage that yields optimistic estimates. Scaling should be fit inside each cross-validation fold, typically within a pipeline.

Question 11Machine Learning Fundamentals and Model Evaluation

A binary classifier's decision threshold is lowered from 0.5 to 0.3. What is the expected effect on precision and recall, all else equal?

  1. Precision increases and recall typically decreases
  2. Both precision and recall increase
  3. Both precision and recall remain unchanged
  4. Recall increases and precision typically decreases Correct answer.

Explanation

Lowering the threshold classifies more cases as positive, which captures more true positives but also admits more false positives. Recall therefore rises while precision generally falls, tracing a movement along the precision-recall trade-off curve.

Question 12Machine Learning Fundamentals and Model Evaluation

Which statement about the area under the ROC curve (AUC) is accurate?

  1. It equals the probability that a randomly chosen positive is scored above a randomly chosen negative Correct answer.
  2. It measures the average squared error of predicted probabilities
  3. It is only defined when classes are perfectly balanced
  4. It depends on the specific decision threshold chosen

Explanation

ROC AUC has a rank-based interpretation: the chance that a random positive instance receives a higher score than a random negative one. Because it integrates over all thresholds, it is threshold-independent, and it remains defined under class imbalance though it can look optimistic there.

Question 13Machine Learning Fundamentals and Model Evaluation

A model achieves ROC AUC of 0.92 but the positive class is rare and the precision-recall curve is poor. What is the best interpretation?

  1. The dataset must be perfectly balanced
  2. The AUC calculation must be incorrect
  3. Ranking is good overall, but the model produces many false positives at practical thresholds Correct answer.
  4. The model has excellent precision at every threshold

Explanation

ROC AUC can remain high under heavy imbalance because the large negative class dominates the false positive rate denominator. The precision-recall curve, which is sensitive to the absolute number of false positives, exposes the practical weakness.

Question 14Machine Learning Fundamentals and Model Evaluation

In a regression problem, which metric is most sensitive to large outliers in the error distribution?

  1. R-squared computed on ranked targets
  2. Mean squared error Correct answer.
  3. Mean absolute error
  4. Median absolute error

Explanation

Squaring errors gives disproportionate weight to large residuals, so mean squared error reacts strongly to outliers. Mean absolute error and median absolute error grow only linearly and are therefore more robust.

Question 15Machine Learning Fundamentals and Model Evaluation

Which practice best guards against selecting a model that merely memorizes the training data?

  1. Removing the validation split to use more training rows
  2. Evaluating on held-out data and comparing against a simpler baseline Correct answer.
  3. Reporting training accuracy as the primary metric
  4. Increasing model depth until training error reaches zero

Explanation

Held-out evaluation plus a baseline comparison separates genuine signal from memorization, because a complex model must beat a simple reference on unseen data to justify itself. Driving training error to zero and dropping validation both remove the evidence needed to detect overfitting.

Question 16Machine Learning Fundamentals and Model Evaluation

Stratified k-fold cross-validation differs from plain k-fold in which way?

  1. It uses a different random seed in every repetition
  2. It preserves the class proportion of the full dataset within each fold Correct answer.
  3. It requires the target to be continuous
  4. It holds out two folds instead of one

Explanation

Stratification keeps each fold's class distribution close to that of the overall dataset, which stabilizes estimates when classes are imbalanced or folds are small. Plain k-fold samples folds without regard to class labels.

Question 17Machine Learning Fundamentals and Model Evaluation

A confusion matrix for a model shows 40 true positives, 10 false negatives, 20 false positives, and 930 true negatives. What is the precision?

  1. 0.67 Correct answer.
  2. 0.40
  3. 0.98
  4. 0.80

Explanation

Precision is TP / (TP + FP) = 40 / (40 + 20) = 0.667. Recall would be 40 / (40 + 10) = 0.80, and accuracy would be (40 + 930) / 1000 = 0.97, so the other values correspond to different metrics.

Question 18Machine Learning Fundamentals and Model Evaluation

Which situation most clearly calls for nested cross-validation?

  1. Performing hyperparameter search and wanting an unbiased estimate of the tuned model's performance Correct answer.
  2. Splitting a dataset once into train and test sets
  3. Computing a confusion matrix at a fixed threshold
  4. Comparing training loss to validation loss during a single fit

Explanation

Nested cross-validation uses an inner loop for hyperparameter selection and an outer loop for performance estimation, preventing the selection process from contaminating the reported score. A single split or a single confusion matrix does not involve tuning inside evaluation.

Question 19Machine Learning Fundamentals and Model Evaluation

A model's Brier score is 0.04 on a test set. How should this be interpreted?

  1. Predicted probabilities are close to observed outcomes on average, since lower Brier scores indicate better calibration and accuracy Correct answer.
  2. The model's ROC AUC must be exactly 0.04
  3. The model's recall is 0.04
  4. The model misclassifies 4% of records by definition

Explanation

The Brier score is the mean squared difference between predicted probabilities and binary outcomes, so values near zero indicate well-calibrated, accurate probability estimates. It is not a classification error rate, not a ranking metric, and not a recall value.

Question 20Machine Learning Fundamentals and Model Evaluation

During 5-fold cross-validation a model shows fold scores of 0.91, 0.90, 0.92, 0.45, and 0.90. What is the most informative next step?

  1. Report the mean of 0.816 as a reliable estimate and stop
  2. Discard the 0.45 fold because low scores are outliers
  3. Investigate the anomalous fold for distribution shift, leakage, or a small or unusual subset Correct answer.
  4. Increase k to 50 to make the anomaly disappear

Explanation

A single sharply deviating fold often signals heterogeneous data, a small or unusual fold, or leakage rather than random noise. Averaging hides this and simply dropping the fold biases the estimate upward, so the cause should be diagnosed before reporting.

Question 21Machine Learning Fundamentals and Model Evaluation

A team trains a model on data that includes a feature recorded only after the outcome occurs. What evaluation risk does this create?

  1. Underfitting, because the feature adds noise
  2. Target leakage, producing optimistic scores that will not hold in deployment Correct answer.
  3. Higher variance in the test metric only
  4. No risk, since more features always improve generalization

Explanation

A feature that is unavailable at prediction time, or that encodes the outcome, constitutes target leakage. The model exploits it during training and evaluation, yielding scores that collapse when the feature is absent in production.

Question 22Machine Learning Fundamentals and Model Evaluation

Two classifiers are compared on the same test set: model A has recall 0.85 and precision 0.40, model B has recall 0.60 and precision 0.75. Which consideration is most relevant to choosing between them?

  1. Which model was trained for more epochs
  2. Which model has the higher training accuracy
  3. The relative business cost of false positives versus false negatives at the intended operating point Correct answer.
  4. Which model has more parameters

Explanation

The two models occupy different points on the precision-recall trade-off, so the decision should follow the relative cost of the two error types in the deployment context. Training accuracy, epochs, and parameter count do not measure the operational consequences of errors.

Question 23Machine Learning Fundamentals and Model Evaluation

A regression model reports R-squared of 0.85 on the test set. Which statement is correct?

  1. The model has 85% accuracy in classifying target values
  2. The model's residuals average exactly 0.85
  3. The model's predictions are 85% likely to be within one unit of the truth
  4. The model explains about 85% of the variance in the target relative to predicting the mean Correct answer.

Explanation

R-squared compares the residual sum of squares to the total variance around the target mean, so 0.85 means 85% of that variance is accounted for. It is not a probability of closeness, not a classification accuracy, and not the mean residual.

Question 24Machine Learning Fundamentals and Model Evaluation

A model's training and validation losses both remain high and nearly equal after many epochs. What does this pattern suggest?

  1. Underfitting, suggesting a need for greater model capacity or better features Correct answer.
  2. Data leakage in the validation set
  3. That the validation split is invalid
  4. Overfitting, requiring stronger regularization

Explanation

When both training and validation error plateau at a high level with little gap, the model lacks the capacity or information to fit the underlying pattern, which is underfitting. Overfitting instead shows a wide gap with low training loss.

Question 25Machine Learning Fundamentals and Model Evaluation

Which statement about leave-one-out cross-validation (LOOCV) is correct?

  1. It cannot be applied to classification problems
  2. It is identical in cost to a single train-test split
  3. It trains one model per observation, which can be computationally expensive for large datasets Correct answer.
  4. It uses a single fixed validation fold regardless of dataset size

Explanation

LOOCV holds out one observation at a time, so the number of fits equals the number of rows, making it costly on large data. It is a valid scheme for classification as well as regression, and it is far more expensive than a single split.

Question 26Machine Learning Fundamentals and Model Evaluation

A model is being evaluated for deployment, and the team wants a single metric that summarizes the trade-off between precision and recall across all possible decision thresholds. Which metric fits this purpose?

  1. Cohen's kappa at the default 0.5 threshold
  2. Accuracy at the default 0.5 threshold
  3. The F1 score at one hand-picked threshold
  4. Average precision, computed as the area under the precision-recall curve Correct answer.

Explanation

Average precision integrates precision over the recall range, so it summarizes the full precision-recall trade-off rather than a single operating point. It is therefore the standard summary when threshold-independent precision-recall behavior matters. Accuracy, F1, and kappa all depend on one fixed threshold.

Question 27Machine Learning Fundamentals and Model Evaluation

A regression model predicts house prices. The test set contains a few homes priced far above the rest, and the team wants an error metric that penalizes these large misses most strongly. Which metric should they report?

  1. Mean absolute percentage error (MAPE)
  2. Median absolute error
  3. Mean absolute error (MAE)
  4. Root mean squared error (RMSE) Correct answer.

Explanation

Squaring the residuals before averaging gives large errors disproportionate weight, so RMSE reacts most strongly to a few extreme misses. MAE and median absolute error grow only linearly, and MAPE is a relative measure that can even shrink for large targets.

Question 28Machine Learning Fundamentals and Model Evaluation

A team evaluates a classifier with 5-fold cross-validation and notices that one fold was built from a single customer whose many transactions all landed in that fold, while the model's other folds contain transactions from different customers. What is the most appropriate fix?

  1. Switch to leave-one-out cross-validation to maximize training data
  2. Increase the number of folds to 10 so each fold is smaller
  3. Use grouped cross-validation so all records from one customer stay in the same fold Correct answer.
  4. Randomize the row order before splitting into folds

Explanation

Records from the same entity are not independent, so splitting them across folds leaks information and inflates validation scores. Grouping by customer keeps each entity wholly within one fold, giving an honest estimate of generalization to unseen customers. Changing fold count or row order does not remove the leakage.

Question 29Machine Learning Fundamentals and Model Evaluation

A team wants to know whether a trained model's predictions on held-out data are better than chance in a statistically meaningful way, and they have only one test set of 200 examples. Which approach is most appropriate?

  1. Compute a confidence interval or run a significance test on the test-set performance Correct answer.
  2. Report the single test-set accuracy and treat it as exact
  3. Retrain the model on the test set and compare training accuracy
  4. Report the training-set accuracy as a proxy for test performance

Explanation

A single point estimate on a finite test set carries sampling uncertainty, so a confidence interval or significance test quantifies whether the result is distinguishable from chance. This respects the test set's role as an unbiased estimate while acknowledging variance. Training on the test set or reporting training accuracy would invalidate the estimate.

Question 30Machine Learning Fundamentals and Model Evaluation

A team compares two models using 10-fold cross-validation on the same folds and finds model A's mean score higher than model B's by a small margin, but the per-fold differences vary widely in sign. What is the most defensible conclusion?

  1. Model B should be chosen because its variance across folds is likely larger
  2. Model A is reliably better because its mean cross-validation score is higher
  3. The difference is not clearly established, so a paired test across folds or more data is needed before preferring A Correct answer.
  4. Both models should be discarded because cross-validation cannot compare models

Explanation

When per-fold differences change sign, the mean gap is small relative to fold-to-fold variability, so the ranking is not yet trustworthy. A paired comparison across the shared folds, or additional data, tests whether the difference exceeds noise. Mean score alone ignores that variability.

Question 31H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

In an H2O-3 cluster, what is the role of the flatfile that is written when a node joins?

  1. It holds the REST API session tokens for authenticated users
  2. It stores the in-memory frame data shared across nodes
  3. It records cluster membership details such as node IP and port so peers can discover each other Correct answer.
  4. It persists the trained model artifacts to disk for later scoring

Explanation

H2O-3 nodes communicate over a multi-node cluster and use a flatfile to advertise membership information (IP, port, and node identity) so other nodes can discover and connect to each other. Frame data and model artifacts are not stored in the flatfile; frames live in distributed memory and models are saved separately.

Question 32H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which H2O-3 component exposes the REST API that clients such as Python and R use to submit work to the cluster?

  1. The Steam service
  2. The H2O Flow notebook only
  3. The REST API server embedded in each H2O node Correct answer.
  4. The Hadoop YARN ResourceManager

Explanation

Each H2O-3 node runs an embedded REST API server; client libraries (Python, R, and others) send requests to that endpoint, which dispatches work to the cluster. Flow is one web client on top of that API, not the API itself, and Steam is a separate deployment-management tool.

Question 33H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

In H2O-3, what does the term 'frame' refer to?

  1. A single JVM thread used for model training
  2. A configuration block inside the flatfile
  3. A distributed, column-typed dataset held across the cluster's memory Correct answer.
  4. A serialized model file on local disk

Explanation

An H2O Frame is the platform's core distributed data structure, analogous to a data frame but partitioned across cluster nodes and column-typed. It is not a thread, a model file, or a flatfile configuration block.

Question 34H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

A data scientist needs to import a large CSV into an H2O-3 cluster so that parsing is distributed. Which call accomplishes this?

  1. h2o.clusterStatus(path)
  2. h2o.downloadCSV(path)
  3. h2o.importFile(path) from the Python client Correct answer.
  4. pandas.read_csv(path) followed by local processing

Explanation

h2o.importFile instructs the cluster to read and parse the file in a distributed manner, producing an H2O Frame. pandas.read_csv builds a local client-side object, while downloadCSV and clusterStatus serve unrelated purposes.

Question 35H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which statement best describes Driverless AI's deployment model compared with H2O-3?

  1. Driverless AI is a fully open-source library that runs only inside a Python process
  2. Driverless AI is only a REST client for an H2O-3 cluster and has no modeling engine of its own
  3. Driverless AI requires a Hadoop cluster and cannot run on a single machine
  4. Driverless AI is a packaged application that automates feature engineering, model tuning, and explanation, typically run as a container or appliance Correct answer.

Explanation

Driverless AI is delivered as a packaged application (commonly a Docker image or native installer) that automates feature engineering, algorithm selection, tuning, and interpretability. It has its own modeling engine and can run on a single machine or larger hardware, so it is not merely an H2O-3 client.

Question 36H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which H2O-3 client call retrieves a model object by its key so it can be used for prediction?

  1. h2o.getModel(model_id) Correct answer.
  2. h2o.init()
  3. h2o.ls()
  4. h2o.removeFrame(model_id)

Explanation

h2o.getModel fetches a model from the cluster by its identifier, returning an object that supports predict. The other calls remove a frame, list keys, or start a connection to a cluster.

Question 37H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

What is the primary purpose of H2O Steam in an H2O deployment?

  1. To store raw training data in a distributed file system
  2. To provide a web-based environment for deploying and managing predictive applications built from H2O models Correct answer.
  3. To train deep learning models on GPUs
  4. To replace the H2O-3 REST API with a binary protocol

Explanation

Steam is a deployment-management layer that lets teams publish, version, and serve H2O models as applications through a web interface. It does not train models, replace the REST protocol, or act as a storage system.

Question 38H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

When starting an H2O-3 cluster from Python, what does h2o.init(nthreads=-1) request?

  1. Use one thread per node
  2. Disable multithreading entirely
  3. Limit the cluster to a single node
  4. Use all available cores on each node Correct answer.

Explanation

In H2O-3, nthreads=-1 means use all available cores on each node for computation. It does not restrict the cluster to one node, disable threading, or force one thread per node.

Question 39H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

An H2O-3 model was trained in a cluster and the analyst wants to score new data in a separate environment without the original cluster. Which approach is appropriate?

  1. Copy the flatfile to the new environment and start a node
  2. Re-run h2o.init() in the new environment and call predict on the old model object reference
  3. Export the training frame as CSV and retrain from scratch each time
  4. Download the model with h2o.download_model or save it as MOJO and load it in the target environment Correct answer.

Explanation

H2O models can be persisted with h2o.save_model (binary) or exported as a MOJO for portable, dependency-light scoring. A model object reference is only valid in the live cluster session, and copying a flatfile or retraining each time does not transfer the trained model.

Question 40H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

What is the key operational difference between an H2O binary model and a MOJO?

  1. A MOJO is a portable, low-dependency scoring artifact, while a binary model requires a compatible H2O runtime to load Correct answer.
  2. A MOJO can be retrained, while a binary model cannot
  3. A binary model is portable across environments, while a MOJO only works inside the original cluster
  4. A binary model is always smaller than a MOJO

Explanation

A MOJO packages the model for scoring with minimal dependencies and is designed for portability across environments and languages. A binary model is the native H2O serialization that requires a compatible H2O runtime to load for further use.

Question 41H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

In a multi-node H2O-3 cluster, how is distributed training work coordinated?

  1. Only the client process performs training, using nodes purely for storage
  2. Each node trains an independent model and the client averages predictions
  3. A leader node coordinates the distributed algorithm while worker nodes process their data partitions Correct answer.
  4. Training is serialized so only one node is active at a time

Explanation

H2O-3 uses a distributed algorithm in which nodes cooperate, with coordination across the cluster and each node processing its local data partition. The client submits the job but does not perform the training math, and nodes do not train fully independent models that are later averaged.

Question 42H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

What does the H2O-3 cluster health endpoint report when a node has become unreachable?

  1. The cluster immediately deletes all frames without warning
  2. The cluster status reflects the missing node and the cluster may enter a degraded or locked state until membership is resolved Correct answer.
  3. The REST API silently redirects all requests to the client machine
  4. The flatfile is automatically rewritten to a single-node configuration

Explanation

Cluster status exposes membership and health; when a node becomes unreachable, the cluster reflects the change and may become degraded or locked while membership is reconciled. Frames are not silently deleted, requests are not redirected to the client, and the flatfile is not auto-converted to single-node.

Question 43H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which H2O-3 feature allows a user to run an AutoML process that trains and ranks multiple algorithms?

  1. h2o.merge()
  2. h2o.splitFrame()
  3. h2o.grid() alone with a single algorithm
  4. h2o.automl() Correct answer.

Explanation

h2o.automl runs an automated pipeline that trains and ranks several algorithms and their variants, producing a leaderboard. h2o.grid tunes one algorithm, splitFrame partitions data, and merge joins frames.

Question 44H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

A team wants to serve Driverless AI model predictions to a production application with low latency. Which artifact is designed for that purpose?

  1. The Driverless AI experiment log file
  2. The raw feature engineering recipes as plain text
  3. The experiment's leaderboard HTML page
  4. The exported scoring pipeline (MOJO) plus the provided scoring server or Python/R scorer Correct answer.

Explanation

Driverless AI can export a scoring pipeline as a MOJO and provides a scoring server or Python/R scoring artifacts for low-latency serving. Log files, recipe text, and leaderboard pages are documentation, not serving artifacts.

Question 45H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

In Driverless AI, what is the role of an experiment's 'accuracy, time, and interpretability' settings?

  1. They control the storage format of the input dataset
  2. They guide the trade-off among model accuracy, training duration, and model interpretability during the automated run Correct answer.
  3. They set the number of cluster nodes required
  4. They determine the REST API authentication method

Explanation

Driverless AI's experiment settings let the user steer the trade-off among accuracy, time budget, and interpretability, which influences feature engineering, model complexity, and tuning. They do not configure cluster nodes, authentication, or storage formats.

Question 46H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which statement correctly describes how H2O-3 handles categorical variables during model training?

  1. It converts every categorical column to a numeric ordinal code chosen by the user
  2. It requires categorical columns to be one-hot encoded by the client before import
  3. It rejects any dataset containing categorical columns
  4. It treats categorical columns as enumerated types that algorithms handle natively, with encoding decided by the algorithm Correct answer.

Explanation

H2O-3 represents categorical columns as enumerated types, and individual algorithms decide how to use them (for example, native categorical splits in tree models). Client-side one-hot encoding is not mandatory.

Question 47H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

What is the purpose of the H2O-3 'leaderboard' produced by AutoML?

  1. It ranks trained models by a chosen metric so the best performer can be selected Correct answer.
  2. It records REST API request timings
  3. It lists the cluster nodes and their memory usage
  4. It stores the raw input frames for auditing

Explanation

The AutoML leaderboard ranks the trained models by a specified metric, helping the user choose the best model. Node and memory information comes from cluster status, and the leaderboard is not a data store or request log.

Question 48H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

An administrator wants H2O-3 to persist models across cluster restarts so they can be reloaded later. Which practice is correct?

  1. Save models to durable storage with h2o.save_model or export MOJOs, then reload them after restart Correct answer.
  2. Delete the cluster and recreate it; models are recovered from the flatfile automatically
  3. Rely on the cluster to keep models in memory indefinitely
  4. Copy the flatfile to back up the models

Explanation

In-memory models do not survive a cluster restart, so they must be saved to durable storage (native save or MOJO export) and reloaded afterward. The flatfile holds membership info and cannot restore models.

Question 49H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

Which H2O-3 configuration is required to run H2O on a Hadoop YARN cluster?

  1. A standalone JVM on the client machine only
  2. A Docker Swarm overlay network
  3. The h2o-yarn application that requests containers from YARN and starts H2O nodes inside them Correct answer.
  4. A Kubernetes ingress controller

Explanation

H2O provides a YARN integration (h2o-yarn) that requests YARN containers and launches H2O nodes within them, forming a cluster. Standalone JVMs, Docker Swarm, and Kubernetes ingress are unrelated to the YARN integration.

Question 50H2O Platform Architecture and Deployment (H2O-3, Driverless AI)

A model trained with H2O-3 uses a column that is renamed in the production scoring data. What is the most likely outcome when scoring with the saved model?

  1. Scoring fails or produces incorrect results because the model expects the original column names Correct answer.
  2. The model automatically retrains on the new column name
  3. The MOJO rewrites the input schema to match
  4. H2O silently renames the column and scores correctly

Explanation

H2O models bind to the column names seen at training time; if a required column is missing or renamed, scoring will fail or misalign. The platform does not silently rename columns, retrain, or rewrite the input schema.