I85PREP
⌘ K
Dataiku ML Practitioner study materials
Dataiku ML Practitioner study resources

Dataiku ML Practitioner practice questions

One question at a time. Build your Dataiku ML Practitioner 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 1Visual machine learning

In a Dataiku visual analysis, which component is used to read a dataset, apply a sequence of transformation steps, and write the result to an output dataset?

  1. A code environment
  2. A scenario trigger
  3. A visual recipe Correct answer.
  4. A project variable

Explanation

A visual recipe is the container that binds input datasets to output datasets and holds an ordered list of transformation steps configured through the graphical interface. Project variables, scenario triggers, and code environments serve unrelated purposes in project configuration and automation.

Question 2Visual machine learning

Which Dataiku visual processor should be used to create a new column whose value is derived row by row from two existing numeric columns?

  1. Join
  2. Group
  3. Formula Correct answer.
  4. Distinct

Explanation

The Formula processor evaluates an expression per row and can reference existing columns, which is exactly the row-level derivation described. Join, Group, and Distinct are relational operations that change row cardinality or combine datasets rather than compute a per-row column.

Question 3Visual machine learning

A practitioner wants to inspect how a single column is distributed before modeling. Which Dataiku feature provides charts and summary statistics directly on a dataset without building a recipe?

  1. The Flow's recipe list
  2. The dataset's Charts tab Correct answer.
  3. The scenario run history
  4. The project's API keys page

Explanation

The Charts tab attached to a dataset lets you build univariate and bivariate visualizations and view summary statistics for columns without creating any recipe. The other listed pages concern orchestration, access, and pipeline structure rather than exploratory statistics.

Question 4Visual machine learning

In the Dataiku visual machine learning interface, what is the purpose of the Design stage of a visual analysis?

  1. To choose the target variable, input features, and training settings before fitting models Correct answer.
  2. To schedule periodic retraining scenarios
  3. To monitor production drift on incoming data
  4. To deploy the trained model as a scored API endpoint

Explanation

The Design stage is where you configure the modeling problem: selecting the prediction target, the features, the train/test split, and the algorithms to train. Deployment, monitoring, and scheduling happen after training and are handled by other parts of the platform.

Question 5Visual machine learning

Which Dataiku model evaluation view lets you compare several trained models side by side on the same metrics?

  1. The Flow's dataset schema view
  2. The code notebook editor
  3. The Models list in the Result tab Correct answer.
  4. The project's permissions screen

Explanation

After training, the Result tab lists every trained model with its performance metrics so you can rank and compare them on a common basis. Schema views, permission screens, and notebooks do not aggregate model metrics.

Question 6Visual machine learning

A visual analysis is configured to predict a numeric revenue value. Which task type should be selected?

  1. Time series forecasting
  2. Binary classification
  3. Clustering
  4. Regression Correct answer.

Explanation

When the target is a continuous numeric quantity, the correct supervised task is regression, which fits a model that predicts real values. Clustering is unsupervised, binary classification requires a two-class target, and forecasting requires a time-ordered setup.

Question 7Visual machine learning

Which Dataiku capability automatically tries multiple algorithms and preprocessing combinations to find well-performing models with limited manual configuration?

  1. The Prepare recipe's sampling setting
  2. The project's wiki
  3. The Sync recipe
  4. AutoML in the visual analysis Correct answer.

Explanation

AutoML explores candidate algorithms and settings and trains a set of models so the practitioner can review and select among them. Sync recipes copy data, sampling settings limit rows processed, and the wiki stores documentation.

Question 8Visual machine learning

In a Dataiku visual analysis, what does the train/test split setting control?

  1. The list of features excluded from the model
  2. The frequency at which the model is retrained
  3. The number of CPU cores allocated to training
  4. The proportion of rows held out to estimate generalization performance Correct answer.

Explanation

The train/test split determines how rows are partitioned so that a model is fitted on one portion and evaluated on unseen rows, giving an estimate of generalization. Core allocation, feature exclusion, and retraining schedules are configured elsewhere.

Question 9Visual machine learning

A practitioner wants to prevent a high-cardinality identifier column from being used as a feature in a visual analysis. What is the most direct action?

  1. Increase the number of trees in the model
  2. Reject the column in the feature handling settings Correct answer.
  3. Change the train/test split ratio
  4. Switch the evaluation metric

Explanation

Feature handling settings let you mark columns as rejected so they are excluded from training, which is the direct way to remove an identifier. Model hyperparameters, split ratios, and metrics do not control which columns enter the model.

Question 10Visual machine learning

In the Dataiku visual analysis Results view, what does the variable importance chart show?

  1. The runtime of each recipe in the Flow
  2. The percentage of missing values per column
  3. The correlation between the target and the row index
  4. The relative contribution of each input feature to the trained model's predictions Correct answer.

Explanation

Variable importance ranks features by how much they influence the fitted model's predictions, helping you understand and prune inputs. Missing-value rates, index correlations, and recipe runtimes are separate diagnostics.

Question 11Visual machine learning

A binary classification model must catch as many positive cases as possible even at the cost of more false positives. Which evaluation view should guide threshold selection?

  1. The Flow's connection name
  2. The dataset's column count
  3. The confusion matrix at candidate thresholds Correct answer.
  4. The model's training duration

Explanation

The confusion matrix computed across candidate thresholds shows the trade-off between true positives and false positives, which is what you need to pick a threshold favoring recall. Training duration, column count, and connection names carry no information about classification trade-offs.

Question 12Visual machine learning

In a Dataiku visual analysis, which setting determines whether categorical features are one-hot encoded, hashed, or otherwise transformed before training?

  1. The scenario's run frequency
  2. Feature handling and encoding settings Correct answer.
  3. The project's language setting
  4. The output dataset's storage format

Explanation

Encoding of categorical variables is controlled through the feature handling and preprocessing settings of the analysis, where you choose how categories are represented for the algorithm. Storage format, scheduling, and interface language do not affect feature encoding.

Question 13Visual machine learning

A model performs extremely well on the training data but poorly on the test set. Which conclusion is most appropriate?

  1. The dataset should be joined to itself
  2. The model is overfitting and should be regularized or simplified Correct answer.
  3. The test set is too large and should be removed
  4. The target variable should be dropped

Explanation

A large gap between training and test performance is the classic signature of overfitting, so reducing model complexity or adding regularization is the appropriate response. Removing the test set, dropping the target, or self-joining the data would not address generalization.

Question 14Visual machine learning

What is the role of the Dataiku Flow in a project?

  1. It stores only the trained model artifacts
  2. It replaces the need for any code recipes
  3. It defines user permissions for the project
  4. It visually represents datasets and recipes and the dependencies between them Correct answer.

Explanation

The Flow is the graphical pipeline view in which datasets and recipes appear as nodes connected by dependency links, making lineage explicit. It is not a model store, it does not forbid code recipes, and permissions are managed elsewhere.

Question 15Visual machine learning

A practitioner wants to evaluate a trained model on a completely separate dataset that was never used during training. Which Dataiku approach supports this?

  1. Delete the test split and rerun training
  2. Increase the number of features in the design
  3. Score the new dataset with the saved model and compare predictions to known labels Correct answer.
  4. Retrain the model on the new dataset and read the training metrics

Explanation

Scoring an independent labeled dataset with the saved model produces predictions that can be compared against ground truth, giving an unbiased performance estimate. Retraining on that data, deleting the split, or adding features does not provide an out-of-sample evaluation.

Question 16Visual machine learning

In a visual analysis for a multiclass target, which metric set is most informative for judging overall performance across all classes?

  1. Only the mean of the target column
  2. Only the training wall-clock time
  3. Per-class precision, recall, and a macro-averaged score Correct answer.
  4. Only the number of rows in the dataset

Explanation

Multiclass problems require looking at precision and recall for each class plus an aggregate such as a macro average, because a single class can dominate overall accuracy. Target means, row counts, and runtimes do not describe classification quality across classes.

Question 17Visual machine learning

A practitioner notices that a categorical feature has thousands of distinct values and the model's test performance is unstable. Which visual analysis action is most reasonable?

  1. Switch the target to a different column
  2. Reduce cardinality by grouping rare categories or reject the feature Correct answer.
  3. Convert the task to clustering
  4. Disable the test set

Explanation

High-cardinality categorical features can cause unstable fits, so grouping infrequent levels or excluding the feature is a standard remedy. Changing the target, removing the test set, or switching to unsupervised clustering avoids rather than fixes the issue.

Question 18Visual machine learning

What does the Dataiku visual analysis do when you enable the option to compute a baseline model?

  1. It deletes models that perform poorly
  2. It deploys the model to production immediately
  3. It trains a simple reference model whose performance the main models should beat Correct answer.
  4. It generates synthetic training rows

Explanation

A baseline provides a simple reference point, such as a majority-class or mean predictor, so you can judge whether a complex model adds real value. It is not a deployment action, it does not remove models, and it does not fabricate data.

Question 19Visual machine learning

A team wants to reuse the exact preprocessing and model from a visual analysis on new data arriving daily. Which Dataiku artifact should they create?

  1. A new chart on the training dataset
  2. A second copy of the training dataset
  3. An additional exploratory notebook
  4. A saved model deployed to the Flow for scoring Correct answer.

Explanation

Saving the model and deploying it into the Flow creates a scoring recipe that applies the same trained pipeline to new datasets. Charts, notebooks, and dataset copies do not operationalize the model for repeated scoring.

Question 20Visual machine learning

In a visual analysis, a practitioner sets the sample settings to use a subset of rows during design. What is the main benefit?

  1. Guaranteed higher accuracy on the test set
  2. Faster iteration on model design while keeping the full dataset for final training Correct answer.
  3. Permanent reduction of the output dataset
  4. Automatic removal of all outliers

Explanation

Sampling during design shortens training cycles so you can explore algorithms and settings quickly, and the analysis can later be retrained on the full data. It does not guarantee accuracy, remove outliers, or permanently shrink outputs.

Question 21Visual machine learning

A visual analysis predicts churn, and the business cares most about not missing churners. Which metric should be prioritized when comparing models?

  1. R-squared
  2. Overall accuracy
  3. Mean absolute error
  4. Recall for the churn class Correct answer.

Explanation

Recall on the positive class measures the share of actual churners that the model catches, which aligns with the goal of minimizing missed churners. Overall accuracy can look strong while missing the minority class, and MAE and R-squared are regression metrics.

Question 22Visual machine learning

A practitioner compares two models: model A has higher test accuracy but its precision on the positive class is much lower than model B's. The use case penalizes false positives heavily. Which model should be preferred and why?

  1. Model B, because higher precision means fewer false positives among predicted positives Correct answer.
  2. Model A, because lower precision indicates a better-calibrated model
  3. Model A, because test accuracy is always the deciding metric
  4. Model B, because precision equals overall accuracy in every case

Explanation

When false positives are costly, precision on the positive class is the relevant signal because it measures the share of predicted positives that are truly positive. Higher overall accuracy does not protect against costly false positives, and precision is not equivalent to accuracy.

Question 23Visual machine learning

During model design, a feature is found to be derived from the target after the fact, such as a cancellation timestamp in a churn model. What is the correct handling in the visual analysis?

  1. Reject the leaking feature and retrain the analysis Correct answer.
  2. Keep it, since higher accuracy is always desirable
  3. Remove the target variable
  4. Convert the task to regression

Explanation

A feature that encodes information only available after the outcome is target leakage; it inflates offline metrics and fails in production, so it must be excluded and the model retrained. Keeping it, changing the task, or dropping the target does not resolve the leakage.

Question 24Visual machine learning

A visual analysis shows strong cross-validated performance, but after deployment the live model's performance degrades steadily over several weeks. Which explanation is most consistent with this pattern?

  1. Distribution drift in the input data relative to the training period Correct answer.
  2. Too many features were rejected during design
  3. The train/test split was set too high
  4. A coding error in the evaluation metric

Explanation

Gradual degradation after a model was validated suggests the relationship between inputs and target has shifted, which is distribution drift and calls for monitoring and retraining. Metric bugs, feature rejection, and split ratios would not produce a steady post-deployment decline.

Question 25Visual machine learning

A team needs to explain individual predictions of a trained visual model to business reviewers. Which Dataiku capability is designed for this?

  1. The Flow's recipe dependency graph
  2. The dataset's row count summary
  3. The project's API key list
  4. Individual prediction explanations showing per-feature contributions Correct answer.

Explanation

Per-prediction explanation views attribute a score to the contributing features, which supports case-level justification for reviewers. Row counts, dependency graphs, and API keys describe structure and access rather than reasons for a prediction.

Question 26Visual machine learning

In a Dataiku visual analysis, which object is trained and stored as the reusable output of the modeling session?

  1. A dashboard tile showing metrics
  2. A saved model with its preprocessing and trained estimator Correct answer.
  3. A code recipe written in Python
  4. A dataset schema snapshot

Explanation

A visual analysis produces a saved model that bundles the selected algorithm, its hyperparameters, and the preprocessing steps learned during training. This saved model can later be deployed, retrained, or scored against new data. Recipes and dashboards are separate Flow objects rather than the modeling output.

Question 27Visual machine learning

A practitioner is configuring a visual analysis and wants the target column to be treated as a two-class label. Which Dataiku prediction type must be chosen?

  1. Clustering
  2. Binary classification Correct answer.
  3. Time series forecasting
  4. Regression

Explanation

Binary classification is the prediction type used when the target has exactly two possible outcomes, such as yes/no or churn/retain. Regression is for continuous numeric targets, clustering is unsupervised, and forecasting predicts future values over time.

Question 28Visual machine learning

In the Dataiku visual analysis Design tab, what is the purpose of the hyperparameter settings panel for a selected algorithm?

  1. It chooses the deployment infrastructure
  2. It controls the algorithm's internal learning settings, such as tree depth or regularization Correct answer.
  3. It defines which columns are excluded from the target
  4. It sets the row sampling ratio for the training set

Explanation

The hyperparameter panel exposes algorithm-specific settings that govern how the model learns, such as number of trees, maximum depth, or regularization strength. These values influence model capacity and overfitting behavior. Column exclusion, sampling, and deployment are handled elsewhere in the interface.

Question 29Visual machine learning

A practitioner wants to quickly gauge how much better a trained model is than random guessing on a binary target. Which metric directly expresses this?

  1. Silhouette score
  2. Mean absolute error
  3. AUC of the ROC curve Correct answer.
  4. Root mean squared error

Explanation

The AUC of the ROC curve measures the probability that a randomly chosen positive is ranked above a randomly chosen negative, with 0.5 representing random guessing. It is therefore a direct read on discrimination above chance. MAE and RMSE are regression metrics, and silhouette measures cluster separation.

Question 30Visual machine learning

In a Dataiku visual analysis, which view lets a practitioner inspect how the model's predicted scores are distributed across actual positive and negative cases?

  1. The Flow graph
  2. The charts tab of the dataset
  3. The model's ROC curve and confusion matrix in the Results view Correct answer.
  4. The project's version history

Explanation

The Results view of a trained model contains diagnostic outputs such as the ROC curve, confusion matrix, and score distributions that separate actual positives from negatives. These let the practitioner see how well scores discriminate the classes. Flow graphs and version history describe project structure rather than model behavior.

Question 31Visual machine learning

A practitioner wants the visual analysis to hold out a portion of rows so performance can be estimated on data the model did not learn from. Which setting accomplishes this?

  1. The deployment mode
  2. The algorithm selection
  3. The train/test split ratio in the design settings Correct answer.
  4. The target column definition

Explanation

The train/test split ratio determines what fraction of rows are reserved as a holdout for evaluating the model. The model trains on the remaining rows and is scored on the holdout, giving an estimate of generalization. Other settings define the learning problem or the estimator rather than the evaluation partition.

Question 32Visual machine learning

Which statement best describes the relationship between a visual analysis and the Flow in a Dataiku project?

  1. A visual analysis cannot access datasets outside its own project
  2. A visual analysis replaces the Flow for modeling tasks
  3. A visual analysis is created from a dataset in the Flow and its saved model becomes a Flow object Correct answer.
  4. A visual analysis writes results only to dashboards

Explanation

A visual analysis is launched from an input dataset that lives in the Flow, and the resulting saved model is itself a Flow object that can feed downstream scoring recipes. This keeps modeling integrated with the project's data pipeline. It does not replace the Flow or bypass project datasets.

Question 33Visual machine learning

When a practitioner opens a saved model in Dataiku and reviews the reported performance metrics, what do those metrics describe?

  1. The speed of the scoring recipe
  2. The number of features used
  3. Performance on the training data only
  4. Performance estimated on the held-out evaluation data Correct answer.

Explanation

The metrics shown for a saved model reflect evaluation on the holdout or cross-validation folds, not the training rows. This is why they are used as an estimate of generalization. Training-set performance and operational details like scoring speed are reported separately.

Question 34Visual machine learning

A practitioner wants to compare several candidate algorithms side by side within one visual analysis session. Which Dataiku feature supports this?

  1. Adding multiple models to the same analysis session and comparing their metrics Correct answer.
  2. Deleting the session between runs
  3. Creating a separate project for each algorithm
  4. Exporting each model to a notebook

Explanation

A single visual analysis session can hold several trained models, each with its own algorithm and settings, and their metrics can be compared directly. This avoids fragmenting experiments across projects. Notebook export and session deletion do not provide side-by-side comparison.

Question 35Visual machine learning

A practitioner wants to reduce overfitting in a gradient-boosted tree model trained in a visual analysis. Which adjustment is most directly aimed at that goal?

  1. Reducing tree depth and increasing regularization Correct answer.
  2. Increasing the number of boosting rounds without limit
  3. Removing the train/test split
  4. Adding the row identifier as a feature

Explanation

Shallower trees and stronger regularization constrain model complexity, which is the standard way to reduce overfitting in boosted trees. Unlimited boosting rounds and identifier features tend to increase overfitting, and removing the holdout eliminates the ability to detect it.

Question 36Visual machine learning

A visual analysis predicts loan default, and the business requires that when the model flags a default it is usually correct. Which metric should be prioritized?

  1. Overall accuracy
  2. Precision on the default class Correct answer.
  3. R-squared
  4. Recall on the default class

Explanation

Precision on the positive class measures how often a positive prediction is actually positive, which matches the requirement that flagged defaults be trustworthy. Recall would favor catching more defaults even at the cost of false alarms. Accuracy can be misleading with imbalanced classes, and R-squared is a regression metric.

Question 37Visual machine learning

A practitioner wants to understand which input columns contributed most to a trained tree model's predictions. Which Results view element is designed for this?

  1. The ROC curve
  2. The variable importance chart Correct answer.
  3. The score distribution histogram
  4. The confusion matrix

Explanation

The variable importance chart ranks features by how much they contribute to the model's predictive decisions. It helps identify influential columns and guide feature engineering. ROC curves, confusion matrices, and score histograms describe overall performance rather than per-feature contribution.

Question 38Visual machine learning

A practitioner wants to try several preprocessing variants for a categorical feature and automatically keep the best-performing combination. Which Dataiku capability is intended for this?

  1. Using the Flow's version control
  2. Writing a custom scoring recipe
  3. The optimization or auto-ML capability that searches algorithm and preprocessing combinations Correct answer.
  4. Manual trial and error in separate sessions

Explanation

Dataiku provides an optimization capability that systematically searches across algorithms and preprocessing settings and retains the best-performing configuration. This replaces slow manual trial and error. Custom scoring recipes and version control serve different purposes.

Question 39Visual machine learning

A practitioner suspects a feature leaks information about the target because it was recorded after the outcome. What is the most appropriate action in the visual analysis?

  1. Keep the feature because it improves test scores
  2. Increase the number of trees to compensate
  3. Lower the train/test split ratio
  4. Remove the feature from the inputs and retrain Correct answer.

Explanation

A feature recorded after the outcome would not be available at prediction time and can inflate apparent performance, so it should be excluded and the model retrained. Retaining it produces a model that cannot be used in production. Adjusting tree count or split ratio does not fix leakage.

Question 40Visual machine learning

A visual analysis is trained on data from one region, and the practitioner wants to know whether the model generalizes to another region. Which approach is most appropriate?

  1. Retrain on the other region and report training accuracy
  2. Evaluate the trained model on the other region's data as an external test set Correct answer.
  3. Remove the target column
  4. Increase the number of features

Explanation

Applying the trained model to a separate dataset from the other region gives an honest estimate of cross-region generalization. Retraining and reporting training accuracy measures fit rather than transfer. Feature changes and target removal do not answer the generalization question.

Question 41Visual machine learning

A practitioner notices the model's test AUC is strong but the confusion matrix at the default threshold shows very few predicted positives. What is the most likely explanation?

  1. The target column is numeric
  2. The decision threshold is too high for the chosen use case Correct answer.
  3. The model is underfitting severely
  4. The dataset has no negative cases

Explanation

A strong AUC indicates good ranking ability, but the threshold used to convert scores into class labels determines how many positives are predicted. A high threshold yields few positive predictions even when discrimination is good. Underfitting would show weak AUC, and a numeric target would indicate a regression task.

Question 42Visual machine learning

A practitioner wants to reduce training time during exploration while still getting a reasonable performance estimate. Which approach is most sensible in the visual analysis sample settings?

  1. Use a sampled subset of rows during design and validate the final model on more data Correct answer.
  2. Disable all metrics
  3. Remove the target column
  4. Train on the full dataset for every experiment

Explanation

Sampling rows during design speeds up experimentation while preserving a reasonable performance signal, and the final candidate can be retrained on more data for accuracy. Training on the full dataset for every experiment is slow, and removing the target or metrics prevents evaluation.

Question 43Visual machine learning

A practitioner has two models: model A has higher test AUC, but model B has a better calibrated score distribution for risk pricing. Which consideration matters most when choosing between them?

  1. Pick the model whose probability estimates match the business use case Correct answer.
  2. Pick the model with more features
  3. Always pick the higher AUC regardless of calibration
  4. Pick the model trained on fewer rows

Explanation

When predicted probabilities are used directly, as in risk pricing, calibration matters as much as ranking ability, so the model whose probabilities are trustworthy for the use case should be preferred. AUC alone does not guarantee calibrated probabilities. Feature count and row count are not the deciding criteria.

Question 44Visual machine learning

A visual analysis shows that a continuous feature has a highly skewed distribution. Which preprocessing step in the analysis is most likely to help a linear model?

  1. Removing the train/test split
  2. Adding the row identifier as a feature
  3. Applying a log or other monotonic transformation to the feature Correct answer.
  4. Increasing the number of classes

Explanation

Monotonic transformations such as a log reduce skew and can make a linear model's relationship with the target closer to linear, improving fit. Removing the holdout, adding identifiers, or changing class counts does not address skew.

Question 45Visual machine learning

A practitioner wants to confirm that a model's strong performance is not an artifact of a particular random split. Which action is most appropriate?

  1. Remove the target column
  2. Use cross-validation to evaluate performance across multiple folds Correct answer.
  3. Rely on a single train/test split
  4. Report only training accuracy

Explanation

Cross-validation evaluates the model across multiple folds, giving a more stable estimate that is less sensitive to a lucky or unlucky split. A single split can be misleading. Training accuracy and target removal do not test split robustness.

Question 46Visual machine learning

A model trained in a visual analysis performs well offline but degrades after deployment. Which factor is most likely responsible when the production data distribution differs from training data?

  1. The choice of AUC as a metric
  2. The presence of a train/test split
  3. Using a saved model instead of a notebook
  4. Data drift between training and production inputs Correct answer.

Explanation

When production inputs drift away from the distribution seen during training, the learned relationships no longer hold and performance degrades. This is a data drift problem rather than an artifact of saved models, splitting, or metric choice. Monitoring input distributions helps detect it.

Question 47Visual machine learning

A practitioner wants to understand why a specific loan application received a high default score. Which Dataiku capability is most directly suited to explaining an individual prediction?

  1. Individual prediction explanation features that attribute contributions to input features Correct answer.
  2. The project's permissions settings
  3. The dataset's row count
  4. The Flow's version history

Explanation

Dataiku offers explanation capabilities that attribute an individual prediction to the contributions of its input features, which is exactly what is needed to justify a single decision. Version history, row counts, and permissions do not explain predictions.

Question 48Visual machine learning

A visual analysis targets a multiclass outcome with imbalanced classes. Which evaluation approach is most informative?

  1. R-squared
  2. Per-class metrics such as precision and recall, plus a macro or weighted average Correct answer.
  3. Overall accuracy alone
  4. The number of trees in the model

Explanation

With imbalanced multiclass targets, overall accuracy can be dominated by the majority class, so per-class precision and recall together with a macro or weighted average reveal how well each class is handled. R-squared is for regression, and tree count is a model setting rather than an evaluation approach.

Question 49Visual machine learning

A practitioner observes that two highly correlated features both appear important, and removing one barely changes performance. What does this suggest about the model?

  1. The train/test split is invalid
  2. The importance is shared between redundant features, so neither alone is essential Correct answer.
  3. The model is severely underfit
  4. The target column is misconfigured

Explanation

When features are highly correlated, their predictive signal is redundant, so importance can be split between them and removing one has little effect. This is a known interpretation caveat for importance charts. It does not by itself indicate underfitting, target misconfiguration, or an invalid split.

Question 50Visual machine learning

A practitioner wants to deploy a visual analysis model so it scores new data automatically each day. Which Dataiku mechanism is most appropriate?

  1. Manually re-running the analysis each morning
  2. Deleting the model after each run
  3. A saved model deployed with a scoring recipe in the Flow, scheduled to run Correct answer.
  4. Exporting the model to a spreadsheet

Explanation

Deploying the saved model and using a scoring recipe within the Flow, then scheduling that recipe, automates daily scoring on new data. Manual re-runs are error-prone, and exporting to a spreadsheet or deleting the model does not provide automated scoring.