DP-100 Certified Data Scientist Practice Exams

Greylock Security Services is a private security contractor started by Evan Rhodes and the company uses Microsoft Azure for its infrastructure, and you are consulting on a project that needs prepared datasets, and an intern asks why training data is required, which statement best describes the role of training data in a machine learning project?

• ❏ A. A reserved portion of data used to evaluate the model after it is trained

• ❏ B. The entire dataset applied at once to train the model

• ❏ C. A subset of records used to train the model so it can learn patterns from examples

• ❏ D. A separate subset used to tune hyperparameters and guide model selection

Aerix Components is a mid sized electronics manufacturer based in Harbor Point San Diego and led by Lena Ortiz. Lena is training a regression model in a notebook to predict how many customers will visit Aerix Boutique stores per day. She wants to log a single metric with MLflow at the end of each training epoch to monitor model performance. Which metric should Lena record in MLflow?

• ❏ A. Mean absolute error

• ❏ B. Accuracy

• ❏ C. Root mean squared error

• ❏ D. Recall

A data engineering team at Meridian Insights plans to perform interactive data cleaning and exploration using a distributed Spark setup in an Azure Machine Learning workspace and they need to know which compute offerings can power their notebooks? (Choose 2)

• ❏ A. Notebook VM

• ❏ B. Google Cloud Dataproc

• ❏ C. Serverless Spark Compute

• ❏ D. Synapse Analytics Spark Pool

• ❏ E. AML Compute Cluster

A regional agritech company called GreenHarvest is using its cloud machine learning workspace to develop a conventional model that will predict which plots are most suitable for different seed varieties. Which machine learning framework is the best fit for this traditional classification task?

• ❏ A. PyTorch

• ❏ B. ONNX

• ❏ C. scikit-learn

• ❏ D. TensorFlow

Scenario: Bistro Solace, a high end Brooklyn restaurant, was founded by Anna Pierce and Marcus Lee who are improving their operations and have adopted Microsoft Azure for analytics. The team is building a training pipeline for a regression model using a dataset of many numeric features that exist on different ranges and Anna needs the numeric features scaled relative to each feature’s minimum and maximum values. Which module should Anna add to the pipeline to perform this min max scaling transformation?

• ❏ A. Cloud Dataflow

• ❏ B. Select Columns in Dataset

• ❏ C. Clean Missing Data

• ❏ D. Normalize Data

Aurora Forecasting is a predictive analytics startup led by Maya Patel and they are preparing an hourly time series dataset that spans about 18 months in Azure Machine Learning Studio and they need to split the records into training and testing sets using the Split Data module while preserving temporal order; which splitting mode should they select?

• ❏ A. Regular Expression Split

• ❏ B. Relative Expression Split mode

• ❏ C. Split Rows with Randomized option enabled

• ❏ D. Recommender Split

A data science team at NorthStar Analytics registered a tabular dataset named ‘model_train_set’ and assigned it to a variable for use by an estimator when running a training script. They want the script to have access to the dataset during the job run. Which estimator property should be configured to provide the training script with the dataset?

• ❏ A. data_reference = model_train_set

• ❏ B. script_params = {“–data”:model_train_set}

• ❏ C. inputs = [model_train_set.as_named_input(“model_train_set”)]

• ❏ D. source_directory = model_train_set

Scenario The Velvet Room is an upscale nightclub in New Metro that also serves as a front for a small business owner. You are hired as a contractor to advise the IT team on machine learning workflows in Microsoft Azure. The team has a dataset with over 180 features. The lead developer plans to train a Two-Class Support Vector Machine binary classifier. The requirement is to compute feature importance with the Permutation Feature Importance module in Azure Machine Learning Designer. The developer lists these actions. a. Upload a dataset to the pipeline. b. Add a Split Data module to create training and test subsets. c. Add a Two-Class Support Vector Machine module to define the SVM estimator. d. Add a Train Model module to produce a trained model. e. Add a Permutation Feature Importance module and attach the trained model and test data. f. Set the performance metric to Classification Accuracy and run the experiment. What is the correct order of these steps?

• ❏ A. c then a then d then b then e then f

• ❏ B. b then c then a then d then e then f

• ❏ C. a then b then c then d then e then f

• ❏ D. a then c then b then d then e then f

Fairlearn is a Python toolkit used to assess models and surface disparities in predictions and performance for specified sensitive attributes, and it can upload dashboard metrics to a StratusML workspace for team review. Which Fairlearn parity constraint matches the description “Use this constraint with any of the reduction-based mitigation algorithms to restrict the loss for each sensitive feature group in a regression model”?

• ❏ A. Equalized odds

• ❏ B. False-positive rate parity

• ❏ C. Bounded group loss

• ❏ D. Demographic parity

• ❏ E. True positive rate parity

• ❏ F. Error rate parity

A mobility startup called HarborRent fit a linear regression using seven days of past scooter rental counts and temperature data and now asks whether the coefficient of determination R squared must always produce values that are zero or greater?

• ❏ A. False

• ❏ B. True

A data scientist at a retail analytics firm is validating a binary classifier developed in Azure Machine Learning studio, which performance measures should they examine to understand how the classifier behaves on positive and negative cases? (Choose 2)

• ❏ A. Area under the ROC curve

• ❏ B. True positives

• ❏ C. Mean absolute error

• ❏ D. False positives

Velvet Echo is an upscale music lounge in Harbor City that also serves as a covert base for a local syndicate and it stores a well structured CSV file in cloud storage and you are advising the venue’s IT staff on how to load that data efficiently into a Pandas DataFrame for analysis which Azure Machine Learning data object should be created to simplify conversion of the CSV into a Pandas DataFrame?

• ❏ A. A file dataset

• ❏ B. A workspace datastore

• ❏ C. None of the listed choices

• ❏ D. A tabular dataset

Sentinel Security operates a protective logistics firm and Maya leads a group of data scientists who want to standardize a reproducible workspace using Azure CLI v2. Which Azure CLI v2 command should Maya run to create a new custom environment for her team?

• ❏ A. az ml environment update

• ❏ B. az ml environment create

• ❏ C. ml_client.environments.create_or_update

• ❏ D. az ml environment show

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

RestoreWorks is an urban restoration startup that Metro Emergency Authority hired to assist recovery after a string of major infrastructure incidents in Harbor City. The firm is led by CEO Ana Rios and she has engaged you as a machine learning consultant. The team will tune hyperparameters that are all discrete and they require every possible combination of values to be evaluated. Which sampling method should you recommend for their tuning process?

• ❏ A. Sobol sampling

• ❏ B. Grid sampling

• ❏ C. Bayesian optimization sampling

• ❏ D. Random sampling

While preparing a forecasting model for a mid sized ecommerce analytics group at Nova Insights you must identify anomalous records in the dataset. Which visualizations are most helpful for highlighting those anomalies? (Choose 2)

• ❏ A. ROC curve

• ❏ B. Box plot

• ❏ C. Confusion matrix

• ❏ D. Scatter plot

• ❏ E. Venn diagram

Dr. Mira Cole at BrightPath Analytics is tuning a deep neural network and decides to raise the learning rate hyperparameter to accelerate convergence during training. What effect does increasing the learning rate have on the training process?

• ❏ A. Cloud TPU

• ❏ B. Backpropagation applies larger weight updates

• ❏ C. Training uses more samples per mini batch

• ❏ D. The network gains additional hidden layers automatically

A predictive analytics group at Meridian Insights built a regression model and now needs to choose an evaluation metric. Which metric is best described by taking the square root of the mean of the squared differences between predicted and actual values and yielding a result in the same units as the target where a larger gap from the mean absolute error signals greater dispersion among individual errors?

• ❏ A. Coefficient of Determination R2

• ❏ B. Relative Absolute Error RAE

• ❏ C. Root Mean Squared Error RMSE

• ❏ D. Relative Squared Error RSE

Nimbus AI Studio supports many open source machine learning frameworks and libraries. Which framework serves as an open interchange standard for representing trained machine learning models?

• ❏ A. scikit-learn

• ❏ B. PyTorch

• ❏ C. TensorFlow SavedModel

• ❏ D. ONNX

Mountvale Motors in Oregon buys and sells pre owned cars and small trucks and the owner plans to use past sales records to train a model that can estimate a resale price from features like manufacturer model engine displacement and odometer reading. What type of machine learning model should they create with automated machine learning to predict a numeric price?

• ❏ A. Classification

• ❏ B. Time series forecasting

• ❏ C. Clustering

• ❏ D. Regression

Harbor Ledger is a regional news publisher led by Benjamin Carter that has grown from a small operation into a recognized media outlet. Carter has hired you as a data engineering consultant to improve analytics workflows and infrastructure. One active project uses Azure Machine Learning Studio to perform feature engineering on a dataset. The team needs to normalize values so they are placed into a feature column with values grouped into bins. Carter tells the team to use Entropy Minimum Description Length MDL binning mode to achieve this. Does Carter’s instruction satisfy the stated project requirement?

• ❏ A. Yes

• ❏ B. No

Scenario Tailored Threads a boutique apparel company based in Manchester expanded by acquiring a label in Barcelona and the team is integrating systems with Microsoft Azure and they asked you to advise on Azure Machine Learning. The engineering group created an Azure Machine Learning compute target named ComputeAlpha using the STANDARD_D2 virtual machine size yet ComputeAlpha shows zero active nodes. A developer has a ws variable that references the Azure Machine Learning workspace and runs the following code Python from azureml.core.compute import ComputeTarget, AmlCompute from azureml.core.compute_target import ComputeTargetException the_cluster_name = “ComputeAlpha” try: the_cluster = ComputeTarget(workspace=ws,name=”ComputeAlpha”) print(“Step1″) except ComputeTargetException: config = AmlCompute.provisioning_configuration(vm_size=”STANDARD_DS3_v2″,max_nodes=6) the_cluster = ComputeTarget.create(ws,”ComputeAlpha”,config) print(“Step2”) The operations lead to no exception and Alfred wants to know whether experiments configured to use the_cluster will run on ComputeAlpha?

• ❏ A. Only interactive runs from the developer machine will use ComputeAlpha and not batch experiments

• ❏ B. No experiments will run because the cluster currently has zero active nodes

• ❏ C. Yes experiments configured to use the_cluster will automatically run on ComputeAlpha

• ❏ D. Experiments will run only after an administrator scales the cluster to one or more nodes

A data science team at Evergreen Analytics is using a visual modeling tool to perform filter based feature selection for a multi class classification problem and the dataset includes categorical predictors that are strongly associated with the target label. Which statistical scoring technique is most appropriate to identify the best categorical features?

• ❏ A. Spearman rank correlation

• ❏ B. Pearson correlation

• ❏ C. Mutual information

• ❏ D. Chi squared test

Nova Instruments is a Seattle based engineering firm led by Maria Chen and it is expanding quickly which is creating new data science requirements. The team needs to construct an Azure Machine Learning pipeline using Designer that trains a model from a comma separated values file hosted on a website and the pipeline must ingest the CSV into Designer with minimal administrative setup while no dataset has yet been registered for that file. Which module should be added to the Designer pipeline to accomplish this?

• ❏ A. Enter Data Manually

• ❏ B. Import Data

• ❏ C. Register a Dataset

• ❏ D. Convert to CSV

Tavern Oak is an upscale Boston bistro founded by Lena Park and Marco Diaz and the restaurant is adopting Microsoft Azure Machine Learning to streamline its operations. You are consulting on several projects and Marco is preparing a new Azure Machine Learning experiment that will train on a small dataset while Lena wants to avoid paying for a cloud virtual machine. Which compute option should Marco select to minimize cost while still handling the low volume training workload?

• ❏ A. Compute cluster

• ❏ B. Inference cluster

• ❏ C. Local compute

• ❏ D. Compute instance

Fill in the blank in the following sentence in the context of machine learning at a fictional cloud provider called Contoso Cloud. [__] is a type of machine learning where you train a model to predict which category or class an item belongs to. For instance a neighborhood clinic could use patient measurements such as height weight blood pressure and fasting glucose to decide whether a patient is diabetic. Which word or words correctly fill the blank?

• ❏ A. Regression

• ❏ B. Statistics

• ❏ C. BigQuery

• ❏ D. Probability

• ❏ E. Classification

A fintech startup called NovaInsights needs to serve model predictions to its mobile clients with very low latency for live decision making. Which Azure service should the team use to host the trained model so it can deliver real time predictions with minimal delay?

• ❏ A. Azure Functions

• ❏ B. Azure App Service

• ❏ C. Azure Machine Learning Studio

• ❏ D. Azure Kubernetes Service (AKS)

Which compute configuration is most suitable for training an image recognition model for a small robotics firm called NovaBots?

• ❏ A. TPU cluster with high memory and 20 nodes

• ❏ B. Single compute instance with low memory and 3 CPU cores

• ❏ C. Managed GPU cluster with medium memory and 20 nodes

• ❏ D. Single compute instance with high memory and 3 CPU cores

Scenario: Meridian Insights is assisting a retail client with their Microsoft Azure data platform and they plan to apply K-means clustering for customer segmentation. The analytics team must define valid stopping rules for the K-means routine. Which of the following conditions can be used to stop the K-means algorithm? (Choose 3)

• ❏ A. Vertex AI

• ❏ B. A fixed number of iterations is reached

• ❏ C. The average distance among members of clusters increases beyond a threshold

• ❏ D. Centroid positions remain unchanged between updates

• ❏ E. The residual sum of squares drops below a preset threshold

A data science team at Meridian Retail Analytics uses Orion Machine Learning to refine a demand forecasting model and they need to tune model hyperparameters. Which approach will be most effective for automated hyperparameter optimization?

• ❏ A. Grid Search

• ❏ B. Random Search

• ❏ C. Orion AutoML service

• ❏ D. Bayesian Optimization

Meridian Insights is the analytics group at Horizon Tech and it is led by Lena Park and Omar Reyes. They adopted Azure Machine Learning and hired you as a consultant to oversee key projects. Lena is training a classification model and she wants to measure how much each feature influenced a single prediction. Which of the following should she examine?

• ❏ A. Permutation feature importance

• ❏ B. Recall and accuracy metrics

• ❏ C. Local feature attributions

• ❏ D. Dataset level feature importance

• ❏ E. Precision and accuracy metrics

A data scientist at Meridian Insights is preparing a dataset for a predictive model and finds many records with missing fields. When using the Clean Missing Data component in Contoso Machine Learning Studio which choice will remove entire records that contain null values?

• ❏ A. Replace missing numeric entries with the column mean

• ❏ B. Impute missing values using hot deck imputation

• ❏ C. Delete rows that contain any missing value

• ❏ D. Drop the entire feature column

• ❏ E. Substitute a custom placeholder for missing entries

• ❏ F. Replace missing categorical values with the most frequent category

Elena Cruz recently joined Crescent Analytics to lead a pilot machine learning program that will classify organizational units and the team is using Azure Machine Learning Designer to build the model and Elena must choose the most appropriate evaluation metric for a classification task. Which metric should she select to measure the model’s ability to distinguish between different classes?

• ❏ A. Log Loss

• ❏ B. R Squared

• ❏ C. Area Under ROC Curve (AUC)

• ❏ D. Mean Absolute Error MAE

Scenario: Meridian Data, a privately held analytics firm led by CEO Elena Cross, has a valuation above thirty two million dollars and was founded after the Meridian Foundation. Elena has asked for help because her engineering team is adopting Azure Databricks within Microsoft Azure Machine Learning for model development. During a workshop the team is debating how to show the files stored in DBFS inside a Databricks notebook. Several solutions have been proposed and only one will correctly list the files in DBFS. Which approach will list files in DBFS from a Databricks notebook?

• ❏ A. ls %fs /mnt/data-files

• ❏ B. %fs dir /mnt/data-files

• ❏ C. %fs ls /mnt/data-files

• ❏ D. ls /mnt/data-files

A data science team at Summit Analytics is creating an experiment in Azure Machine Learning Studio and needs to partition a dataset into training and holdout subsets. Which module should they use to perform that split?

• ❏ A. Group Data into Bins module

• ❏ B. Split Data module

• ❏ C. Clip Values module

• ❏ D. Group Categorical Values module

Meridian Recruiting LLC is a search firm based in Chicago and it is led by CEO Aisha Grant. The firm stores candidate records as TSV files in an Azure Blob container that is registered as a datastore in an Azure Machine Learning workspace. A data engineer merged all TSV files and registered the combined dataset under the name candidateSet_5 using the Azure Machine Learning Python SDK. The engineer asks whether candidateSet_5 can be converted into a Pandas DataFrame by calling python candidateSet_5.to_pandas_dataframe(). What should you tell them?

• ❏ A. No, candidateSet_5 cannot be converted into a Pandas DataFrame using to_pandas_dataframe() even if it was defined as tabular

• ❏ B. Yes, candidateSet_5 can be converted into a Pandas DataFrame using to_pandas_dataframe() if the dataset was correctly registered as a tabular dataset

Within Microsoft Azure machine learning projects statistics and mathematics form the foundation and it is important to understand the technical vocabulary used by statisticians mathematicians and data scientists. The difference between a predicted label and the observed label can be treated as a measure of error yet observed values come from sampled observations that can show random variation. To make explicit the comparison between a predicted value written as “y-hat” and an observed value y we call the difference between them the . We can then aggregate the across all validation predictions to compute the model loss as a measure of predictive performance. Which word or words correctly complete the sentence?

• ❏ A. Coefficient of determination

• ❏ B. Random sampling variance

• ❏ C. Root Mean Squared Error

• ❏ D. Residuals

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

Scenario: Nordal Materials a manufacturing group based in Stockholm runs production sites across Europe and Asia. An engineer named Priya will train a machine learning model on an Azure virtual machine that already has the needed libraries and tooling installed. The dataset is small to moderate in size and dynamic scaling is not necessary. Which compute targets should Priya choose to best fit these conditions? (Choose 2)

• ❏ A. Compute clusters

• ❏ B. Serverless compute

• ❏ C. Attached compute

• ❏ D. Inference clusters on Azure Kubernetes Service

• ❏ E. Local compute

A boutique firm named Meridian Analytics must build a model to forecast equity prices using data stored in a PostgreSQL server and a workload that needs GPU acceleration. You need to provision a virtual machine image that arrives with common machine learning frameworks and GPU drivers already installed. Which virtual machine image best fits this requirement?

• ❏ A. Deep Learning Virtual Machine Windows edition

• ❏ B. Deep Learning Virtual Machine Linux edition

• ❏ C. Data Science Virtual Machine Windows edition

• ❏ D. Deep Learning VM image on Google Compute Engine

Riverside Analytics was founded by Elena Morales and is currently worth more than forty million dollars. Elena created the Riverside Foundation and she serves as both president and board chair. She has asked you to advise her IT group as they deploy Microsoft Azure Machine Learning. In a planning meeting the team intends to spin up a new cluster inside an Azure Databricks workspace and they want to know what happens behind the scenes when a cluster is created. Which of the following accurately describes the actions that occur when Azure Databricks provisions a new cluster?

• ❏ A. When a workspace is created you are given a reserved group of virtual machines and clusters reuse machines from that reserved group

• ❏ B. Azure Databricks deploys a managed appliance into your subscription and then launches driver and worker virtual machines in the appliance using the VM sizes you choose

• ❏ C. The platform provisions a single dedicated virtual machine to execute all notebooks and jobs for the workspace

• ❏ D. A serverless style compute pool is automatically created and Azure Databricks draws scaled resources from that pool for interactive workloads

Maria Rivera works at Nova Labs and has trained a model using Azure Machine Learning. She needs a deployment that runs every 24 hours to process an input file and save prediction outputs to a designated Azure Blob Storage container. Which deployment approach should she choose?

• ❏ A. Azure Functions

• ❏ B. Online endpoint

• ❏ C. Batch endpoint

• ❏ D. Azure Kubernetes Service

While preparing a customer ledger for BrightData Labs you must drop repeated rows and keep only the final occurrence of each duplicate entry. Which pandas DataFrame method should you call to obtain a DataFrame that preserves only the last instance of every duplicate?

• ❏ A. duplicated(keep=’last’)

• ❏ B. dropdupes(retain=’last’)

• ❏ C. drop_duplicates(keep=’last’)

• ❏ D. drop_duplicates()

Fairlearn is a Python library used to examine disparities in model predictions across protected attributes and it integrates with Acme Machine Learning Studio so engineers can run experiments and upload dashboard metrics to an Acme workspace. The selection of a parity constraint depends on the mitigation method and the fairness goal. Which Fairlearn parity constraint is described by the following sentence? “This constraint can be applied with any mitigation algorithm to reduce differences among protected groups and in a binary classification setting it ensures each group has a similar proportion of false positive predictions”?

• ❏ A. Equalized odds

• ❏ B. False positive rate parity

• ❏ C. Demographic parity

• ❏ D. Error rate parity

• ❏ E. True positive rate parity

• ❏ F. Bounded group loss

Aquila Analytics was founded by Sarah Patel and now runs a large network of industrial sensors and the data science team is tuning hyperparameters with Azure Machine Learning. Which distribution type should Sarah choose to support discrete hyperparameters?

• ❏ A. Uniform

• ❏ B. Categorical

• ❏ C. QNormal

• ❏ D. LogNormal

At DataForge Analytics we want to build a model that estimates a numeric target using past input features and their observed numeric outcomes. What type of model meets this requirement?

• ❏ A. Multinomial classification model

• ❏ B. Vertex AI

• ❏ C. Ordinal classification model

• ❏ D. Log linear regression model

• ❏ E. Binomial classification model

• ❏ F. Regression model

A data science team at Harbor Analytics is comparing visualization methods to evaluate a new binary classifier and they want a chart that highlights model precision during assessment. Which visualization should they use?

• ❏ A. A Receiver Operating Characteristic curve visualization

• ❏ B. A precision recall curve visualization

• ❏ C. A box plot visualization of prediction scores

• ❏ D. A binary classification confusion matrix visualization

A high end supper club named The Velvet Parlor operates as a front for a syndicate and it hired you to advise on machine learning processes. The proprietor Lucien Vale and his staff are training a model with Microsoft Azure Machine Learning Studio and their dataset contains rows with null entries so they plan to use the Clean Missing Data module to detect and address missing values. Which module parameter should they choose to properly handle rows that include missing values?

• ❏ A. Substitute missing values with the most frequent value

• ❏ B. Drop the entire feature column that contains missing values

• ❏ C. Delete rows containing missing values

• ❏ D. Provide a custom replacement value for missing entries

• ❏ E. Replace missing values with the mean of the column

• ❏ F. Substitute missing values with the median of the column

Maya Lopez is a data science consultant at Nexa Insights and she must advise the operations team on a low or no code platform that allows business analysts to train machine learning models without writing code. Which solution should Maya recommend to Nexa Insights?

• ❏ A. Vertex AI AutoML

• ❏ B. Jupyter Notebooks in Azure Machine Learning Studio

• ❏ C. Azure CLI v2

• ❏ D. Azure Automated Machine Learning

Dr Elena Rivers at Beacon Data Lab is retraining a production machine learning model in Vertex AI to keep its predictions accurate and relevant. Which primary action should she prioritize to ensure the updated model will not reduce performance when it replaces the current model?

• ❏ A. Automate data preprocessing with a reproducible pipeline

• ❏ B. Compare outputs of the new model with outputs of the existing model

• ❏ C. Enable Vertex AI continuous evaluation and model monitoring

• ❏ D. Replace the existing model whenever a predefined replacement rule is met

BrightMart data scientists must deploy a trained model to serve live predictions for their online storefront. The deployment must provide very low latency and maintain high request throughput. Which deployment setup should they select for optimal online serving?

• ❏ A. Cloud Functions

• ❏ B. Compute Engine with GPUs

• ❏ C. Cloud Run

• ❏ D. Google Kubernetes Engine with autoscaling

You are advising Sentinel Cyber Labs and helping Maria who leads the IT group with an anomaly detection initiative on Azure Machine Learning. The team is training a support vector machine from scikit-learn and Maria needs to evaluate the trained model’s accuracy on a held out test set. Which scikit-learn method should the team use to obtain the classifier’s accuracy on the test data?

• ❏ A. mlflow.log_metric

• ❏ B. predict

• ❏ C. fit

• ❏ D. score

At Northbridge Healthcare on Solace you are advising Lena on a customer retention initiative and she needs to run counterfactual what if analysis to identify the smallest changes to customer features that would switch a churn prediction to a retained outcome. Which solution will best allow Lena to perform counterfactual analysis and refine retention actions?

• ❏ A. MLflow

• ❏ B. Azure Machine Learning compute

• ❏ C. Responsible AI Dashboard

• ❏ D. ML Pipelines

Azure Automated Machine Learning streamlines many parts of model building but still allows engineers to apply some manual constraints to the pipeline. Which controls can a practitioner use to influence the AutoML workflow? (Choose 2)

• ❏ A. Override the automatic choice of the primary evaluation metric

• ❏ B. Exclude specific model families from the candidate pool

• ❏ C. Disable scaling and normalization of numeric inputs

• ❏ D. Block selected feature transformations and encodings

NovaChem is a UK specialty materials firm with its headquarters in Manchester and operations in many countries and the company employs thousands of staff across several sites. Priya Kumar from the IT team needs to train a classification machine learning model using Azure Machine Learning and she wants to obtain the highest performing model without writing any code. Which development approach should Priya choose to meet this requirement?

• ❏ A. VS Code extensions for Azure ML

• ❏ B. Azure Machine Learning Designer pipelines

• ❏ C. Azure AutoML

• ❏ D. Azure CLI for Machine Learning

A data engineering group has a reusable YAML file that defines an Azure machine learning compute cluster. You need to publish step by step guidance on the intranet so team members can reliably provision the cluster with Azure CLI using the YAML file. Which Azure CLI command should the documentation show?

• ❏ A. az ml resource create -f cluster-config.yml

• ❏ B. az ml compute create -f cluster-config.yml

• ❏ C. az ml compute create –name analyticsCluster –sku Standard_D3s_v2 –min-instances 2 –max-instances 6 –type AmlCompute –resource-group analyticsRg –workspace-name analyticsWorkspace

• ❏ D. az ml compute add -f cluster-config.yml

You are working in a notebook within Azure Machine Learning studio and you must add Python libraries so they apply only to the notebook’s current kernel and do not alter other kernels or environments. Which notebook magic command should you use to install packages into the active kernel?

• ❏ A. %conda

• ❏ B. %load

• ❏ C. %pip

• ❏ D. !pip

The correct answer is A subset of records used to train the model so it can learn patterns from examples.

Training data supplies the examples the learning algorithm uses to adjust model parameters and form patterns that map inputs to outputs. In supervised learning the training set is usually labeled and must be representative of the problem domain so the model can generalize from those examples.

During training the model is fitted on the training set often through multiple passes or batches while other datasets are reserved for tuning and final evaluation. The training set is distinct from sets used to tune hyperparameters or to provide an unbiased final performance estimate.

A reserved portion of data used to evaluate the model after it is trained is incorrect because that phrase describes the test set, which is held back until after training to give an unbiased measure of how the model performs on new data.

The entire dataset applied at once to train the model is incorrect because you typically split data and train iteratively in batches or epochs. Using the entire dataset without separation prevents proper validation and increases the risk of overfitting and misleading results.

A separate subset used to tune hyperparameters and guide model selection is incorrect because that describes the validation set, which is used during development to compare models and choose hyperparameters while the training set is used to fit the model parameters.

When answering dataset role questions look for keywords like train, validation, and test and map them to learning, tuning, and final evaluation respectively.

Root mean squared error is the correct choice for Lena to record in MLflow.

Lena is solving a regression problem because she predicts a numeric count of customers per day and Root mean squared error measures prediction error in the same units as the target. It reports the square root of the average squared error so it penalizes larger mistakes more strongly and that makes it useful for spotting regressions in model quality across epochs.

MLflow can log a single scalar metric per epoch and recording Root mean squared error as the metric lets Lena monitor whether the model is reducing the kinds of large errors that matter for store traffic forecasts.

Mean absolute error is also a regression metric and it reports average absolute deviation in the same units, but it treats all errors equally and does not penalize large errors as strongly as the chosen metric, so it is not the correct option here.

Accuracy measures the proportion of correct class labels and it is not applicable to continuous numeric predictions, so it is not appropriate for this regression task.

Recall measures how many true positives are found by a classifier and it does not apply to predicting numeric counts, so it is not the correct metric to log for this problem.

When a question describes predicting a numeric value focus on regression metrics. Log a scalar such as RMSE each epoch with mlflow.log_metric to track model performance trends.

The correct options are Serverless Spark Compute and Synapse Analytics Spark Pool.

Serverless Spark Compute is a managed, on demand Spark runtime integrated into Azure Machine Learning that lets you run distributed interactive Spark sessions from notebooks without provisioning and managing cluster infrastructure.

Synapse Analytics Spark Pool provides a dedicated Spark runtime in Azure Synapse that can be used for interactive distributed data exploration and cleaning and it can be integrated or accessed from notebooks for large scale Spark workloads.

Notebook VM is incorrect because it refers to a single node notebook environment and not to a managed distributed Spark service. Note that the term Notebook VM is a legacy name and Azure Machine Learning now uses Compute Instances for single node notebooks which still do not provide a Spark cluster.

Google Cloud Dataproc is incorrect because it is a Google Cloud service and not an Azure compute option available inside an Azure Machine Learning workspace.

AML Compute Cluster is incorrect because Azure Machine Learning compute clusters are general purpose compute targets for training and jobs and they do not provide a managed Spark runtime for interactive distributed Spark notebooks.

When you see the phrase distributed Spark look for Azure services that explicitly offer Spark runtimes and rule out single node notebook instances and offerings from other clouds.

The correct option is scikit-learn.

scikit-learn is designed for traditional supervised learning on tabular data and it provides a broad set of well tested classical classification algorithms such as logistic regression, decision trees, random forests, and gradient boosting that are easy to train and evaluate. It has a simple API and is ideal for quick prototyping and productionizing conventional classification models in a cloud machine learning workspace.

PyTorch is a deep learning framework that is excellent for building custom neural networks and research oriented models but it is generally more complex and resource intensive than needed for a standard tabular classification task.

TensorFlow is also focused on deep learning and scalable model training and serving. It can solve classification problems but it is usually a heavier choice compared with libraries that specialize in classical machine learning for tabular data.

ONNX is an open model format and runtime for interoperability and deployment rather than a primary training framework. It is useful for exporting or running models but it is not the best answer when choosing a framework to build a conventional classifier from scratch.

When the question mentions conventional or tabular classification favor scikit-learn for its simple API and classical algorithms that make iteration fast and implementation straightforward.

The correct option is Normalize Data.

The Normalize Data module applies normalization methods such as min max scaling which rescales numeric features to a common range typically 0 to 1. This transformation computes each feature’s minimum and maximum and then scales values relative to those bounds which matches the requirement in the question.

The module offers MinMax normalization in Azure Machine Learning designer and it is the appropriate choice when models require features on the same scale or when you want values expressed relative to each feature’s min and max.

The Cloud Dataflow option is a Google Cloud data processing service and not an Azure Machine Learning designer module. It is not the correct choice for an Azure pipeline task that performs min max scaling.

The Select Columns in Dataset module only selects which features to include or exclude and does not perform any scaling. It is used to shape the dataset but not to change numeric ranges.

The Clean Missing Data module handles null or missing values by imputation or removal and does not perform min max normalization. It is useful for preparing data but it does not match the scaling requirement.

When a question asks for min max scaling choose the module that explicitly mentions normalization and rule out modules that only select columns or handle missing values.

The correct answer is: Relative Expression Split mode.

The Relative Expression Split mode lets you define a condition based on row indices or timestamp values so you can create a contiguous training period and a contiguous testing period which preserves temporal ordering for time series forecasting. Using an expression you can keep the earliest portion of the dataset for training and the most recent portion for testing which prevents lookahead bias and better reflects real forecasting scenarios.

Regular Expression Split is incorrect because regular expressions are used to match and split text based on patterns and they do not provide a natural way to select contiguous rows by time or index to preserve chronological order.

Split Rows with Randomized option enabled is incorrect because enabling randomization shuffles rows and destroys temporal order which can leak future information into the training set and invalidate time series evaluation.

Recommender Split is incorrect because it is tailored for recommender system data partitioning and user item scenarios and it does not provide the straightforward contiguous time based cut needed for general time series forecasting.

When a question asks you to preserve temporal order choose a split that selects contiguous rows by index or timestamp. Prefer using relative expression or explicit time cutoffs and avoid any randomized split for time series.

The correct answer is inputs = [model_train_set.as_named_input(“model_train_set”)].

The estimator inputs = [model_train_set.as_named_input(“model_train_set”)] attaches a dataset consumption configuration to the training job so the dataset is mounted or downloaded and made available to the training script. Using as_named_input on a registered Dataset returns a DatasetConsumptionConfig that the Estimator recognizes in its inputs list and exposes inside the run as a path or mount point for the script to consume.

data_reference = model_train_set is incorrect because Estimator does not use a property named data_reference to pass registered Datasets. DataReference belongs to older APIs and is not the standard way to provide a Dataset to an Estimator.

script_params = {“–data”:model_train_set} is incorrect because script_params only passes literal command line arguments and will not convert a Dataset object into a dataset consumption configuration. If you want the script to receive a path you must pass the Dataset via inputs and then supply the input’s path or name as a script parameter.

source_directory = model_train_set is incorrect because source_directory should point to the folder containing the training script and its dependencies and it is not used to supply data to the run.

When a question asks which estimator property provides a dataset to the training script remember that you must create a DatasetConsumptionConfig with as_named_input and pass it via the estimator inputs property. Then use a script argument if you need the path inside the script.

a then b then c then d then e then f is correct. This order uploads the dataset first, then creates training and test splits, then defines the Two Class Support Vector Machine estimator, then trains the model, then computes permutation feature importance with the trained model and test data, and finally sets the metric and runs the experiment.

You must upload the dataset before you can split it, so the Upload a dataset step comes first. Splitting the data next ensures you have a proper training split to feed into the Train Model module. Defining the Two Class Support Vector Machine before training gives the Train Model module an estimator to fit. Train Model then produces a trained model that the Permutation Feature Importance module can use when it is supplied with the test split. The final step is to choose the performance metric such as Classification Accuracy and run the experiment to evaluate and report the feature importance results.

c then a then d then b then e then f is wrong because it attempts to add the Two Class Support Vector Machine before the dataset is uploaded and it tries to train before the data has been split into training and test subsets. The Train Model would not have the proper training split as input in that sequence.

b then c then a then d then e then f is wrong because it places the Split Data and estimator steps before the dataset upload. Splitting or configuring modules before the data exists means the modules will not have the dataset inputs they require.

a then c then b then d then e then f is wrong because it defines the estimator before creating the training and test splits. That ordering risks not providing the Train Model module with the correct training split when wiring the pipeline and evaluating feature importance.

When you see workflow ordering questions imagine the data flow and the inputs each module needs. Upload the data first and make sure the training split exists before you run Train Model or compute feature importance.

The correct option is Bounded group loss.

Bounded group loss is a parity constraint intended for use with reduction based mitigation algorithms to limit the loss experienced by each sensitive feature group in regression problems. It is implemented in Fairlearn’s reductions module so you can apply reduction based methods to enforce a maximum group loss while optimizing overall predictive performance.

Equalized odds is not correct because it refers to a classification requirement that true positive rates and false positive rates be equal across groups rather than bounding per group loss for regression.

False-positive rate parity is not correct because it targets equality of false positive rates across groups in classification problems and it does not constrain regression loss for each group.

Demographic parity is not correct because it enforces equal positive outcome rates across groups in classification and it does not implement a reduction based per group loss bound for regression.

True positive rate parity is not correct because it focuses on equal true positive rates across groups in classification rather than bounding group losses in regression.

Error rate parity is not correct because it aims for equal overall error rates across groups in classification scenarios and it does not describe the reduction based bounded loss constraint used for regression.

When a question mentions regression and reduction-based mitigation think of constraints that bound loss per group and look for bounded group loss rather than rate based parity definitions.

The correct answer is False.

R squared is defined as one minus the sum of squared errors divided by the total sum of squares and it compares model error to the error of predicting the mean. If the model’s predictions are worse than always predicting the mean then the sum of squared errors can be larger than the total sum of squares. That yields one minus SSE over SST that is negative so R squared can be negative.

Negative values often arise when you compute R squared on new data and the model generalizes poorly or when the regression omits an intercept or when you use a constrained or non least squares fitting method. By contrast, when you fit ordinary least squares with an intercept on the same data used for training, the intercept only model is a feasible baseline and OLS minimizes SSE so R squared will lie between zero and one in that special situation.

True is incorrect because it asserts that R squared must always be zero or greater. That statement is false for the general definitions and common evaluation scenarios, so the correct choice is False.

Remember that R squared equals one minus SSE divided by SST so if SSE exceeds SST the value will be negative. On exam questions check whether the score is computed on training or test data and whether an intercept is included.

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

True positives and False positives are correct.

True positives count the instances that are actually positive and that the classifier also predicted as positive, so this measure directly shows how the model behaves on positive cases and relates to sensitivity and recall.

False positives count the instances that are actually negative but that the classifier predicted as positive, so this measure directly shows how often the model mistakenly labels negative cases as positive and relates to specificity and precision trade offs.

Area under the ROC curve is not the best answer for this question because it is an aggregate, threshold independent measure of ranking performance rather than a direct count of how the model classifies positive and negative cases.

Mean absolute error is incorrect because it is a regression error metric that measures average absolute difference between predicted and actual numeric values and it does not describe classification behavior on positive and negative cases.

When a question asks about behavior on positive and negative cases focus on confusion matrix counts such as true positives and false positives rather than on aggregate or regression metrics.

The correct option is A tabular dataset.

A tabular dataset represents structured, row and column data such as a well formed CSV and it provides parsing and schema inference so you can work with the data as a table. It also exposes a to_pandas_dataframe method that lets you load the dataset directly into a Pandas DataFrame for analysis, which is the simplest and most efficient path in Azure Machine Learning for this use case.

Using a tabular dataset also lets you register the structured dataset with the workspace so analysts and experiments can consistently access the same parsed table without reimplementing parsing logic.

A file dataset is intended for collections of files or unstructured data and it does not provide the same table semantics or the direct to_pandas_dataframe convenience that a tabular dataset provides. You would need to read and parse the CSV yourself if you used a file dataset.

A workspace datastore is a storage endpoint abstraction that provides a connection to blob or file storage and it is not itself a dataset. You must create a dataset that references a datastore before you get the table behaviors needed to load into Pandas.

None of the listed choices is incorrect because Azure Machine Learning includes a Tabular Dataset option that directly supports loading CSV into a Pandas DataFrame, so a valid choice is available.

When the question asks about loading CSV into Pandas in Azure ML look for the dataset type that models rows and columns and remember that TabularDataset includes the to_pandas_dataframe method to perform the conversion.

The correct option is az ml environment create.

az ml environment create is the Azure CLI v2 command that registers a new custom environment in your Azure Machine Learning workspace. Environments capture dependency specifications and runtime configuration such as conda files and container images so teams can reproduce experiments and deployments consistently.

az ml environment update is used to modify an existing environment definition and it does not create a new environment. Use update only when you already have a registered environment that you want to change.

ml_client.environments.create_or_update is a Python SDK method from the azure AI ML library and not an Azure CLI v2 command. It can create or update environments when you are writing Python code but it is not the CLI command requested in the question.

az ml environment show returns the details of an existing environment and it does not create or register a new environment. Use show when you want to inspect a registered environment.

When you need to register a reproducible environment run the CLI command az ml environment create with a YAML or conda specification file. If you are scripting in Python then use the SDK method ml_client.environments.create_or_update instead.

Grid sampling is the correct choice because the team must evaluate every possible combination of discrete hyperparameter values.

Grid sampling enumerates the Cartesian product of each discrete hyperparameter domain and therefore will try every combination when you run the full grid. That property makes it the right method for an exhaustive search of a discrete and finite search space.

Grid sampling does become expensive as the number of hyperparameters and the number of values per hyperparameter grow because the total experiments increase multiplicatively. You should confirm the total number of combinations is feasible and plan for parallel execution or pruning if needed.

Sobol sampling is a low discrepancy quasi random sequence method that aims for even coverage of continuous spaces so it does not guarantee that every discrete combination will be evaluated and is therefore not suitable here.

Bayesian optimization sampling builds a model of the objective and selects promising points adaptively so it focuses on finding good configurations with fewer evaluations and will not exhaustively evaluate all combinations.

Random sampling picks configurations at random and can cover spaces efficiently for some problems but it does not systematically enumerate every combination and will miss many discrete combinations unless an impractically large number of trials is run.

When a question requires you to evaluate every combination pick Grid sampling and compute the total number of combinations first so you can judge the cost.

The correct options are Box plot and Scatter plot.

A Box plot summarizes the distribution of a numeric variable by showing quartiles and whiskers and it marks individual points outside the expected range as outliers which makes anomalous records easy to spot on a single field.

A Scatter plot displays the relationship between two variables and it reveals points that fall away from clusters or expected trends so it is useful for finding anomalies that involve interactions between fields.

ROC curve evaluates classifier performance across thresholds by plotting true positive rate against false positive rate and it is not designed to highlight individual anomalous records in a dataset.

Confusion matrix summarizes counts of predicted versus actual classes and it helps assess model errors but it does not show record level outliers or continuous distributions.

Venn diagram illustrates overlaps between sets and membership relations and it is not suitable for identifying numeric outliers or multivariate deviations.

When you need to find anomalies look for visualizations that show distributions or relationships such as box plots for single variables and scatter plots for pairs of variables.

The correct answer is Backpropagation applies larger weight updates.

Raising the learning rate increases the multiplier applied to the gradient during each update so backpropagation produces larger changes to the model weights. Larger updates can speed up convergence when the rate is well chosen and they can also cause overshooting or unstable training if the rate is too large.

Cloud TPU is incorrect because the learning rate does not change the hardware used for training. A Cloud TPU is an accelerator and it is unrelated to how large each gradient step is.

Training uses more samples per mini batch is incorrect because the learning rate scales the gradient and it does not alter the mini batch size or the number of samples processed per step.

The network gains additional hidden layers automatically is incorrect because changing the learning rate does not modify the model architecture. Hidden layers are added only when you redesign or reconfigure the network.

When a question mentions learning rate remember that it directly scales gradient updates and affects update magnitude and stability. Do not confuse learning rate with batch size or hardware options when choosing the correct answer.

The correct option is Root Mean Squared Error RMSE.

Root Mean Squared Error RMSE is computed by taking the square root of the mean of the squared differences between predicted and actual values which yields a result in the same units as the target variable. The squaring step increases the influence of larger errors so a larger gap from the mean absolute error signals greater dispersion among individual errors rather than a simple average magnitude.

Coefficient of Determination R2 measures the proportion of variance explained by the model and is unitless so it does not match the description of taking a square root to return units of the target.

Relative Absolute Error RAE is a ratio that compares absolute error to a baseline absolute error so it is relative and unitless and it does not involve squaring and taking a square root.

Relative Squared Error RSE compares squared error to a baseline squared error and thus is a relative measure that does not take the square root to return values in the target units.

When an option mentions the metric returns values in the same units as the target look for a square root in the formula and remember that RMSE penalizes large errors more than MAE.

The correct option is ONNX.

ONNX is an open interchange format that represents trained machine learning models in a framework agnostic way. It defines a common model representation and operator sets so models can be exported from one framework and imported into another for inference and optimization.

The ONNX ecosystem includes runtimes and tools such as ONNX Runtime that perform cross platform inference and optimizations for models saved in the ONNX format. That broad runtime and tool support is what makes ONNX an interchange standard for trained models.

scikit-learn is a Python library for building and training classical machine learning models and not an interchange specification for saving models across frameworks.

PyTorch is a deep learning framework and not a framework agnostic model interchange standard. PyTorch can export models to ONNX but the framework itself is not the open interchange format.

TensorFlow SavedModel is TensorFlow’s native serialized format for saving models and it is specific to TensorFlow. It is not an open interchange standard that is framework agnostic in the same way that ONNX is.

When a question asks about an interchange or portability standard look for formats that are described as framework agnostic and supported by multiple runtimes. ONNX is a common example to watch for.

The correct option is Regression.

Regression models are designed to predict continuous numeric values such as a car resale price. Automated machine learning for tabular data can use past sales records and features like manufacturer model engine displacement and odometer reading to learn a function that outputs a numeric price estimate.

Classification predicts discrete categories rather than continuous numbers so it is not suitable when the goal is to estimate a numeric price.

Time series forecasting focuses on predicting future values that depend on time and temporal patterns. The question asks to map features to a numeric price for individual vehicles so general regression on tabular data is the correct choice.

Clustering is an unsupervised method that groups similar records and does not directly predict a numeric target value so it will not produce price estimates.

When the target variable is a continuous number think regression. If the target is one of several classes think classification and if the task is about predicting future values over time think time series forecasting.

The correct option is No. Carter’s instruction to use Entropy Minimum Description Length MDL binning does not satisfy the stated requirement.

Entropy Minimum Description Length binning in Azure Machine Learning Studio is a supervised discretization technique that uses the target label and an entropy based criterion to choose cut points that reduce description length. Because it relies on class labels it is intended to create bins that are predictive of the target rather than simply to normalize or group values into a single unsupervised feature column.

For a requirement that simply asks to normalize values and place them into a feature column with values grouped into bins the team should use an unsupervised binning method such as equal width, equal frequency, or k means within the Discretize module. Those approaches do not require target labels and will produce bins based on the feature distribution rather than on label information.

The option Yes is incorrect because choosing Entropy MDL assumes access to and use of the target variable and it produces bins optimized for classification performance rather than for plain normalization or label independent binning.

Read the question for whether the task is supervised or unsupervised and check if the technique requires a target label before selecting a binning method.

Yes experiments configured to use the_cluster will automatically run on ComputeAlpha

The cluster resource exists in the workspace even though it shows zero active nodes. An AmlCompute cluster can be created with a low or zero starting node count and Azure Machine Learning will allocate or scale up nodes when a job is submitted so experiments that target the compute will be scheduled and run once nodes are provisioned up to the cluster’s configured limits.

The code shown calls ComputeTarget(workspace=ws, name=”ComputeAlpha”) so it simply binds to the existing compute rather than recreating it. The provisioning configuration in the except block would only apply if the lookup raised an exception. Jobs may wait briefly while VMs are allocated but they do not require a human to first start nodes.

Only interactive runs from the developer machine will use ComputeAlpha and not batch experiments is incorrect because compute targets are available to both interactive sessions and submitted experiments. Azure Machine Learning does not restrict a compute target only to interactive use.

No experiments will run because the cluster currently has zero active nodes is incorrect because a zero node count does not prevent job execution. The service will request and provision nodes for the cluster when an experiment is submitted, assuming the cluster has a positive max_nodes and the subscription has available quota.

Experiments will run only after an administrator scales the cluster to one or more nodes is incorrect because administrative manual scaling is not required for normal job submission. The platform will automatically scale the cluster up to the configured maximum when work is queued, although an administrator might be needed only to change VM sizes or resolve quota issues.

When you see an AmlCompute cluster with zero nodes remember that Azure Machine Learning can autoscale on job submission. Check the min_nodes and max_nodes settings and confirm subscription quotas before assuming manual intervention is required.

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

The correct option is Chi squared test.

The Chi squared test measures association between categorical predictors and a categorical target by comparing observed and expected counts in a contingency table. It directly tests independence and produces a statistic or p value that can be used to rank categorical features for filter based selection in a multi class classification problem.

The Chi squared test is commonly implemented in visual modeling tools and feature selection libraries because it is simple to compute and interpretable for nominal variables and multiclass labels. It is appropriate when cell counts are sufficiently large and when features and labels are both categorical.

Spearman rank correlation evaluates monotonic relationships and requires ordinal or continuous numeric inputs. It is not designed for nominal categorical predictors and so it is not suitable here.

Pearson correlation measures linear association between two continuous variables and assumes numeric data. It does not apply to nominal categorical predictors and so it is not the right choice.

Mutual information does capture general dependence and can handle categorical variables, but it often requires careful estimation and is less commonly presented as the standard statistical contingency test in many visual filter selection workflows. The exam choice favors the chi squared test because it directly tests independence using contingency tables.

When features and the target are categorical look for tests that work on contingency tables and test for independence such as the chi squared test.

Import Data is the correct module to add to the Designer pipeline.

Import Data allows Designer to read a comma separated values file directly from an HTTP or HTTPS web location and bring it into the pipeline without first registering a dataset. This approach satisfies the requirement for minimal administrative setup because the module performs the ingestion at runtime and does not require precreating a named dataset in the workspace.

Enter Data Manually is intended for small ad hoc tables that users type into the canvas and it is not suitable for importing a full CSV from a website.

Register a Dataset is the action of creating a named dataset in the workspace and it requires the data to be available and registered ahead of time so it does not meet the requirement to ingest a web‑hosted CSV with minimal setup.

Convert to CSV is used to change data format within a pipeline and it does not fetch or import a file from an external web location so it will not accomplish the required ingestion step.

When a question emphasizes minimal administrative setup and a file hosted on a website choose a module that can ingest from a URL directly. The Import Data module is designed for this scenario and avoids the extra step of registering a dataset first.

The correct option is Local compute.

Choosing Local compute allows Marco to run the experiment on his own machine so there are no additional cloud virtual machine charges and it can handle low volume training on a small dataset without provisioning remote compute resources.

Local compute is appropriate when you do development and light training locally and when you want to avoid paying for managed VM or cluster time. It works well for single node, non distributed training and for quick iterations on small datasets.

Compute cluster is not ideal because clusters are meant for scalable and distributed training and they incur cloud costs due to provisioned nodes, so they are overkill for a small local workload.

Inference cluster is not correct because inference clusters are used for serving deployed models and real time or batch predictions rather than for training experiments.

Compute instance is not the best choice here because a compute instance is a cloud hosted development VM and it will generate charges while it is running, which conflicts with Lena’s desire to avoid paying for a cloud virtual machine.

When a question emphasizes avoiding cloud VM costs and the workload is small, prefer a local development or local compute option for training if that is allowed by the scenario.

The correct answer is Classification.

Classification refers to supervised machine learning tasks where the model predicts a discrete category or class for each input. For example a neighborhood clinic can use patient measurements such as height weight blood pressure and fasting glucose to predict whether a patient is diabetic which is a binary class label. Classification models often return class probabilities and then a final class label is chosen by applying a threshold or selecting the highest probability.

Regression is incorrect because regression predicts continuous numeric values such as a blood glucose reading rather than assigning an input to a category.

Statistics is incorrect because it is a broad discipline that underpins data analysis and machine learning rather than a task of predicting categories.

BigQuery is incorrect because it is a cloud data warehousing product and not a type of machine learning task.

Probability is incorrect because it is a concept used within models and not the name of a prediction task that assigns class labels.

Look for words like category or class to identify classification questions. If the problem asks for a numeric value then it is likely a regression problem.

The correct option is Azure Kubernetes Service (AKS).

Azure Kubernetes Service (AKS) is well suited for hosting trained models that must deliver real time predictions with minimal delay because it runs containerized inference servers on dedicated nodes and gives you control over CPU, memory, and GPU resources. It supports horizontal scaling and load balancing so you can maintain low latency under varying load and it integrates easily with common model servers such as TensorFlow Serving and TorchServe.

Azure Functions is a serverless option that can be useful for lightweight or sporadic workloads but it often has cold starts and resource limits that make it a poor fit for heavy models and strict low latency requirements.

Azure App Service is designed for hosting web apps and APIs and it does not provide the same level of container orchestration, pod level autoscaling, or GPU access that you need for production grade, low latency model serving.

Azure Machine Learning Studio is focused on building, training, and experimenting with models and while Azure Machine Learning can orchestrate deployments it typically deploys production real time endpoints onto Azure Kubernetes Service (AKS). The Studio interface itself is not the standalone runtime that guarantees the minimal delay required for live decision making.

When a question asks for minimal latency for real time inference favor container orchestration on dedicated nodes with GPU support and autoscaling rather than generic web hosting or serverless options that can introduce cold starts.

The correct answer is Managed GPU cluster with medium memory and 20 nodes.

This option is best because GPU acceleration is well suited to training convolutional neural networks used in image recognition and a managed cluster provides scaling and operational simplicity. Using a medium memory configuration across multiple nodes lets the team train on reasonable batch sizes and distribute work to reduce wall clock time without taking on the extra complexity of custom distributed infra.

TPU cluster with high memory and 20 nodes is not the best fit because TPUs are typically optimized for very large scale workloads and specific model types and they can add cost and operational constraints that are unnecessary for a small robotics firm. TPUs may also require additional engineering to adapt some training pipelines.

Single compute instance with low memory and 3 CPU cores is insufficient because CPUs with low memory will be too slow and may run out of memory for typical image datasets and modern deep learning models. Training on that instance would be impractically long.

Single compute instance with high memory and 3 CPU cores is also not suitable because high memory does not replace the parallel compute capabilities of GPUs. CPU only training will be much slower and will limit iteration speed during model development.

When choosing compute for model training match the compute type to the workload and think about scale as well. For image recognition prioritize GPU or TPU and prefer managed clusters when you want easier scaling and operations.

A fixed number of iterations is reached, Centroid positions remain unchanged between updates and The residual sum of squares drops below a preset threshold are correct stopping rules for the K-means algorithm.

A fixed number of iterations is reached is a practical termination condition that bounds runtime and guarantees the routine will stop even if perfect convergence is slow or not achieved. It is commonly used as a fallback to limit compute resources.

Centroid positions remain unchanged between updates is the classical convergence criterion for K-means because the algorithm alternates assignments and centroid updates and it has converged when centroids no longer move. Stable centroids imply assignments will not change further and the objective has been reached.

The residual sum of squares drops below a preset threshold monitors the K-means objective which is the sum of squared distances within clusters and stopping when this error is sufficiently small ensures a desired level of cluster compactness. This checks for adequate improvement in the objective rather than arbitrary conditions.

Vertex AI is incorrect because it is the name of a cloud ML service and not a termination condition for the algorithm. It does not describe a stopping rule for K-means.

The average distance among members of clusters increases beyond a threshold is incorrect because K-means aims to decrease within cluster distances and an increase would indicate a worsening fit or instability rather than a valid convergence signal. Stopping rules typically watch for decreases or stabilization rather than increases.

When choosing stopping rules look for options that signal convergence or a change in the algorithm objective function and avoid answers that name products or describe conditions that imply a worsening fit.

The correct answer is Orion AutoML service.

Orion AutoML service is the most effective choice because it is a managed, automated hyperparameter optimization solution that tightly integrates tuning with model training and evaluation. The service can orchestrate parallel trials, apply efficient search strategies under the hood, and pick the best configuration based on validation metrics, so it reduces manual effort and scales well for production workflows.

Grid Search is incorrect because it exhaustively evaluates combinations and quickly becomes infeasible for large or high dimensional search spaces, making it an inefficient choice for automated tuning.

Random Search is incorrect because while it can be simple and sometimes effective, it is not a managed, end to end service and it lacks the adaptive efficiency and orchestration that AutoML systems provide.

Bayesian Optimization is incorrect in this context because it describes an optimization technique rather than a standalone managed service, and teams typically use it as one algorithm inside an AutoML system rather than as the full automated solution itself.

When a question asks for automated and scalable hyperparameter tuning favor a named, managed AutoML service and emphasize its managed and end to end capabilities in your reasoning.

The correct answer is Local feature attributions.

Local feature attributions provide per-feature contribution scores for a single prediction and they show how much each input influenced that specific output. They are implemented by techniques such as SHAP and LIME which assign positive or negative contributions for individual instances and they are the right choice when Lena wants to explain why one example was classified a certain way.

Permutation feature importance measures how much the model performance drops when a feature is shuffled across the dataset and it therefore reflects global importance across many examples rather than the influence on a single prediction.

Dataset level feature importance explicitly describes global importance for the dataset and it does not tell you how features affected one particular prediction.

Recall and accuracy metrics are aggregate performance metrics that describe how well the classifier performs overall and they do not provide per-feature explanations for an individual prediction.

Precision and accuracy metrics are also overall model metrics and they do not indicate the contribution of each feature to a single output.

Carefully note whether the question asks about a single prediction or overall model behavior. If it asks about one prediction then look for methods that provide local explanations such as SHAP or LIME rather than aggregate metrics or dataset level importance.

The correct choice is Delete rows that contain any missing value.

Choosing Delete rows that contain any missing value tells the Clean Missing Data component to remove entire records that contain nulls so any row with a missing entry is dropped from the dataset and only complete rows remain for modeling.

Replace missing numeric entries with the column mean is an imputation option that fills numeric gaps with the column average and it does not remove rows.

Impute missing values using hot deck imputation also fills missing entries by borrowing values from similar records and it preserves rows instead of deleting them.

Drop the entire feature column removes a whole column that has missing values rather than removing the records that contain nulls, so it is not the option that deletes rows.

Substitute a custom placeholder for missing entries replaces missing values with a specified token and it keeps the records intact rather than removing them.

Replace missing categorical values with the most frequent category imputes categorical gaps by using the mode and it does not remove any records.

Read the action word in the option carefully and watch for terms like delete or drop versus replace or impute to determine whether rows are removed or values are filled.

The correct option is Area Under ROC Curve (AUC).

Area Under ROC Curve (AUC) measures how well a classifier ranks positive instances above negative ones by summarizing the ROC curve that plots true positive rate against false positive rate across all thresholds. It returns a value between 0 and 1 where higher values indicate better separability and it is threshold independent so it directly captures the model’s ability to distinguish between classes.

Area Under ROC Curve (AUC) is often preferred when classes are imbalanced or when you care about ranking and discrimination rather than a single decision threshold. Azure Machine Learning Designer reports AUC as a standard classification metric for these reasons.

Log Loss measures the negative log likelihood of predicted probabilities and penalizes confident incorrect predictions. It evaluates probability calibration and average prediction uncertainty rather than the model’s ranking or separability, so it does not directly answer how well the model distinguishes classes.

R Squared is a regression metric that indicates the proportion of variance explained by a model. It applies to continuous targets and is not appropriate for evaluating classification separability.

Mean Absolute Error MAE is also a regression metric that measures average absolute differences between predicted and actual numeric values. It does not measure class discrimination or ranking and so it is not suitable for this classification question.

When a question asks about a metric that measures a model’s ability to separate classes focus on metrics that are threshold independent and favor AUC for ranking and discrimination. Verify whether the task is classification or regression before choosing metrics like MAE or R Squared.

The correct option is %fs ls /mnt/data-files.

This command uses the Databricks notebook file system magic and the ls subcommand to list the contents of the DBFS mount point. The %fs magic is the supported notebook directive for interacting with DBFS and ls is the correct subcommand to display files and directories.

ls %fs /mnt/data-files is incorrect because the magic and the subcommand are reversed. The percent sign must prefix the magic name, so the correct form places %fs first.

%fs dir /mnt/data-files is incorrect because dir is not a valid %fs subcommand for listing DBFS contents. The supported listing operation is ls rather than dir.

ls /mnt/data-files is incorrect because that is a plain shell or language level command and it does not use the Databricks %fs magic to access DBFS directly from a notebook cell.

Remember that Databricks notebook magic commands start with a percent sign and require the exact syntax. For listing DBFS content use %fs ls followed by the path.

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

The correct option is Split Data module.

The Split Data module is designed to partition a dataset into two subsets and it lets you specify the fraction or the number of rows to allocate to each subset. You can choose random sampling and set a random seed for reproducibility and you can also use stratified sampling when you need to preserve class distributions between the training and holdout sets.

The Group Data into Bins module is intended to convert continuous numeric variables into discrete bins for feature engineering. It does not split the dataset into separate training and holdout sets.

The Clip Values module is used to cap or floor numeric values to handle outliers or enforce limits. It modifies values rather than partitioning the data so it will not create training and holdout subsets.

The Group Categorical Values module is used to combine or regroup categories in a categorical feature to reduce cardinality. It is a transformation tool and it does not perform dataset splitting.

When you need to create training and holdout sets look for modules that explicitly mention split or partition and confirm the available sampling options and the ability to set a random seed for reproducible results.

The correct answer is: Yes, candidateSet_5 can be converted into a Pandas DataFrame using to_pandas_dataframe() if the dataset was correctly registered as a tabular dataset.

This is correct because the Azure Machine Learning Python SDK provides a tabular dataset type that implements a to_pandas_dataframe() method. If candidateSet_5 was registered as a TabularDataset when the engineer merged the TSV files then calling to_pandas_dataframe() will return a pandas DataFrame containing the combined rows and columns.

You must confirm that the dataset was created or registered using the tabular dataset creation APIs so that the service recognized the files as a table. If the registration preserved the schema and column parsing for the TSV files then the SDK will parse the text into columns and load it into pandas on demand.

No, candidateSet_5 cannot be converted into a Pandas DataFrame using to_pandas_dataframe() even if it was defined as tabular is incorrect because the method is available for datasets registered as tabular. The only time to_pandas_dataframe() would not work is when the dataset was registered as a file style dataset or when the data was not parsed into a tabular schema, in which case you would need to read the raw TSV files with pandas.read_csv and a tab separator.

When you see questions about converting datasets to pandas remember to check whether the dataset was registered as a TabularDataset. Only tabular datasets expose to_pandas_dataframe() in the Azure Machine Learning Python SDK.

The correct option is Residuals.

Residuals are the per observation differences between the observed value y and the predicted value written as y hat. They represent the error for each prediction and you can aggregate the residuals across a validation set by squaring them and then taking a mean or root mean square to compute loss and performance metrics.

Root Mean Squared Error is not correct because it is an aggregate metric derived from the residuals rather than the single observation difference that fills the blank.

Coefficient of determination is not correct because it refers to R squared which summarizes explained variance across the dataset and does not denote the per prediction difference between y and y hat.

Random sampling variance is not correct because it describes variability due to sampling and not the individual prediction error measured by a residual.

When asked about the difference between a prediction and the observed value look for terms that mean per-observation error and avoid options that name aggregated metrics.

The correct options are Attached compute and Local compute.

Attached compute is appropriate because it lets you use an existing Azure virtual machine or other external VM as a compute target in your workspace. It avoids recreating environments when the VM already has the needed libraries and it is suitable when the dataset is small to moderate and you do not need autoscaling.

Local compute is also appropriate because it runs training on the machine where the job is launched and it is ideal for interactive development, debugging, and short or small scale training runs on a preconfigured VM.

Compute clusters are managed, autoscaling clusters intended for distributed or large scale training. They add overhead and complexity that is unnecessary when you have a single preconfigured VM and no need for dynamic scaling.

Serverless compute provides managed execution and automatic scaling and it may not allow using a specific preinstalled VM environment. It is not the best fit when you want to run training directly on an existing VM with custom tooling.

Inference clusters on Azure Kubernetes Service are designed for serving deployed models and inference workloads rather than for training. They are not the correct compute target for Priya’s training job.

When a VM already has the required tools and you do not need autoscaling choose attached compute or local compute for training and reserve cluster or serverless options for scalable or managed scenarios.

Deep Learning Virtual Machine Linux edition is the correct option.

Deep Learning Virtual Machine Linux edition arrives with common machine learning frameworks like TensorFlow and PyTorch and with NVIDIA GPU drivers and CUDA libraries preinstalled so it is ready for GPU accelerated model training out of the box. The Linux edition is the standard choice for GPU workloads because driver support and compatibility for many open source ML tools are best on Linux.

The image also provides optimized binaries and package management that let the team start training quickly without spending time on low level driver and library installation. That makes it the best fit when you need a Compute Engine image that already includes GPU drivers and popular ML frameworks.

Deep Learning Virtual Machine Windows edition is not the best answer because the Linux edition is the preferred and more widely supported environment for GPU accelerated machine learning and many ML tools and drivers have stronger support on Linux.

Data Science Virtual Machine Windows edition is incorrect because that product name refers to a Microsoft Azure offering and not a Google Cloud image provisioned on Compute Engine.

Deep Learning VM image on Google Compute Engine is incorrect because it is an ambiguous label and the exam expects the specific Linux Deep Learning VM image. The explicit Deep Learning Virtual Machine Linux edition is the image that meets the requirement for preinstalled frameworks and GPU drivers.

When a question asks for a prebuilt image with GPU drivers and ML frameworks choose the Linux Deep Learning VM by default because Linux Deep Learning VMs include CUDA and common frameworks and they are the standard for GPU accelerated training.

Azure Databricks deploys a managed appliance into your subscription and then launches driver and worker virtual machines in the appliance using the VM sizes you choose is correct.

When you create a cluster Azure Databricks provisions managed resources in your subscription and then launches a driver node and one or more worker nodes using the VM sizes you selected. The driver coordinates notebook execution and job orchestration while workers run the parallel tasks. Clusters are therefore created as VMs inside the managed appliance rather than being taken from a preallocated pool or reduced to a single machine.

When a workspace is created you are given a reserved group of virtual machines and clusters reuse machines from that reserved group is incorrect because Databricks does not reserve a fixed set of VMs at workspace creation. Clusters are provisioned on demand in the managed resource group and are not simply reused from a preallocated pool for the workspace.

The platform provisions a single dedicated virtual machine to execute all notebooks and jobs for the workspace is incorrect because workloads run on driver and worker nodes and require isolation and scaling. Even single node clusters still have a driver and do not represent a single VM that handles all workspace activity.

A serverless style compute pool is automatically created and Azure Databricks draws scaled resources from that pool for interactive workloads is incorrect because serverless compute is a distinct offering and it is not implicitly created for every workspace. Interactive clusters are normally provisioned as VMs and serverless endpoints are a separate managed feature.

On the exam remember that Azure Databricks provisions managed resources into your subscription and launches driver and worker VMs per cluster rather than using a reserved VM pool or a single VM for the whole workspace.

Batch endpoint is the correct choice.

Batch endpoint is built for offline and large scale inference where you process files or datasets rather than single low latency requests. A Batch endpoint runs the model as a job on managed compute and can read input files and write prediction outputs directly to an Azure Blob Storage container. You can schedule that job to run every 24 hours by using Azure Machine Learning pipelines or an external scheduler and the Batch endpoint will handle the orchestration and scaling for the periodic processing.

Azure Functions are serverless and suited to small event driven tasks and quick API style work. They have execution time limits and require more manual handling of large model binaries and scaling for heavy inference, so they are not ideal for scheduled bulk file processing.

Online endpoint is intended for low latency real time predictions where clients send individual requests and expect immediate responses. It is not optimized for processing large input files or producing batch output files to blob storage on a schedule.

Azure Kubernetes Service can host custom inference services and provide control over scaling and deployment, but it requires you to manage the cluster and orchestration. That makes it more operational overhead compared with using a managed Batch endpoint that is designed for scheduled batch jobs.

Pay attention to words like every 24 hours and input file. Those indicate a batch processing requirement which points to a batch endpoint rather than an online endpoint or serverless function.

The correct option is drop_duplicates(keep=’last’).

The DataFrame method drop_duplicates(keep=’last’) removes duplicate rows and preserves only the final occurrence of each set of duplicates when you set the keep parameter to ‘last’. You can return a new DataFrame or apply the operation in place and this behavior directly matches the requirement to keep only the last instance of every duplicate.

duplicated(keep=’last’) is incorrect because duplicated returns a boolean mask that marks which rows are duplicates and it does not itself remove rows. You would need to apply that mask with additional indexing to obtain a DataFrame without earlier duplicates.

dropdupes(retain=’last’) is incorrect because there is no such pandas method. The correct pandas method name is drop_duplicates and the parameter names differ from the ones shown.

drop_duplicates() is incorrect because calling drop_duplicates without arguments uses the default keep behavior which is ‘first’. That default will preserve the first occurrence and drop later ones which is the opposite of keeping only the final occurrence.

When you need to keep the final occurrence of duplicates remember to set keep=’last’ on drop_duplicates. Also recall that duplicated returns a boolean mask and does not drop rows by itself.

The correct answer is False positive rate parity.

False positive rate parity specifically targets the rate at which negative instances are incorrectly labeled as positive and it seeks to ensure that this false positive rate is similar across protected groups. This constraint can be applied with many mitigation algorithms to reduce differences among groups and in a binary classification setting it ensures each group has a similar proportion of false positive predictions.

Equalized odds is not correct because it requires both false positive rates and true positive rates to be similar across groups rather than only the false positive rate.

Demographic parity is not correct because it enforces equal overall positive prediction rates across groups regardless of the true label and it does not specifically ensure similar false positive rates.

Error rate parity is not correct because it refers to matching the overall misclassification rate across groups and it mixes false positives and false negatives instead of focusing solely on false positives.

True positive rate parity is not correct because it targets equal sensitivity or the rate of correctly predicted positives across groups rather than the false positive frequency.

Bounded group loss is not correct because it constrains or bounds group losses or risks and it does not specifically describe equalizing false positive rates across groups.

When a question mentions matching the rate of incorrect positive predictions look for the phrase False positive rate or FPR and contrast it with overall positive rates or overall error rates.

The correct option is QNormal.

A QNormal distribution draws from a normal distribution and then quantizes the sampled values to a fixed step size so the outputs are discrete. This makes it suitable when you need integer or stepped numeric hyperparameters while keeping the underlying normal sampling behavior.

Uniform is not correct because a plain uniform distribution describes continuous sampling across a range and does not by itself provide quantized or stepped discrete values.

Categorical is not correct in this case because it represents a set of unordered distinct choices rather than a quantized numeric range. Categorical is appropriate for named options or labels and not for discretized numeric hyperparameters that follow a statistical distribution.

LogNormal is not correct because it is a continuous skewed distribution that produces positive real values and does not provide inherent quantization for discrete parameter values.

When a question mentions discrete or integer hyperparameters look for quantized distributions such as q* variants, and reserve categorical for unordered label choices.

The correct option is Regression model.

A Regression model is the appropriate choice because the task describes predicting a continuous numeric target from past input features and observed numeric outcomes. Regression models are the supervised learning models designed to estimate numeric values rather than discrete categories.

A Regression model can use algorithms such as linear regression, decision tree regression, or more advanced methods to learn the relationship between inputs and a continuous response so you can make numeric predictions on new data.

Multinomial classification model predicts which discrete category out of three or more possible classes applies and it is not used for predicting continuous numeric values.

Vertex AI is a Google Cloud platform for training and deploying models and it is not a type of predictive model itself, so it does not answer the question about which model type to use.

Ordinal classification model predicts ordered discrete categories rather than a continuous numeric target, which means it does not meet the requirement to estimate numeric outcomes.

Log linear regression model refers to a specific transformed or specialized regression formulation in some contexts, but the question asks for the general model type and the expected answer is the broader Regression model, not a specialized variant.

Binomial classification model predicts one of two discrete classes and it is therefore not appropriate for estimating continuous numeric targets.

When a question asks you to predict a numeric value think regression and rule out classification options or product names.

The correct option is A binary classification confusion matrix visualization.

The A binary classification confusion matrix visualization presents the counts of true positives false positives true negatives and false negatives so you can directly compute precision as true positives divided by the sum of true positives and false positives. Viewing the confusion matrix makes the contributions to precision explicit for a chosen decision threshold and it therefore highlights precision during assessment.

A Receiver Operating Characteristic curve visualization plots true positive rate against false positive rate and it focuses on discrimination and ranking rather than the positive predictive value so it does not directly show precision.

A precision recall curve visualization does show precision versus recall across thresholds and it is useful for examining precision behavior across decision thresholds however it emphasizes tradeoffs across thresholds rather than highlighting the precision value at a specific threshold which is what a confusion matrix makes explicit.

A box plot visualization of prediction scores displays the distribution of predicted probabilities or scores and it can help assess separation between classes but it does not provide the confusion counts needed to calculate precision directly so it is not the best choice to highlight precision.

When a question asks which chart highlights precision look for visualizations that expose true positive and false positive counts such as a confusion matrix for a chosen threshold.

All Azure Questions are from my DP-100 Udemy Course and certificationexams.pro

Delete rows containing missing values is correct because the Clean Missing Data module can be configured to remove any record that has one or more missing entries and that parameter explicitly drops those rows.

This choice discards entire observations with nulls and is appropriate when missingness is rare or when removing incomplete rows will not introduce bias or remove too much data. The Clean Missing Data module also offers imputation and column removal options, but the delete rows setting directly addresses rows that include missing values by removing them from the dataset.

Substitute missing values with the most frequent value is an imputation option that fills nulls with the mode of the feature and does not remove rows, so it does not meet the requirement to delete rows that contain missing entries.

Drop the entire feature column that contains missing values removes whole columns rather than rows, and that can discard useful information across all records, so it is not the correct parameter when the goal is to handle rows with missing values.

Provide a custom replacement value for missing entries is another imputation strategy that replaces nulls with a user supplied value and therefore preserves rows instead of deleting them, so it is not the parameter that removes rows with missing data.

Replace missing values with the mean of the column computes and substitutes the column mean for nulls and keeps the rows, so it does not fulfill the requirement to delete rows containing missing values.

Substitute missing values with the median of the column is similar because it imputes nulls with the median and retains the records, so it is not the correct choice when the desired action is to drop rows with missing entries.

Read the question carefully to see whether it asks to remove records or to impute missing values. If the exam asks about handling rows with nulls look for wording like delete rows or remove records rather than imputation terms such as mean, median, or most frequent value.

Azure Automated Machine Learning is the correct choice.

Azure Automated Machine Learning provides a guided, low and no code experience that lets business analysts train, evaluate, and deploy supervised learning models without writing code. It automates model selection, feature handling, and hyperparameter tuning and it integrates with Azure Machine Learning Studio so teams can operationalize models with minimal developer involvement.

Vertex AI AutoML is a Google Cloud service that also offers AutoML capabilities but it is not the Azure service referenced in this question and it would not be the recommended Azure solution.

Jupyter Notebooks in Azure Machine Learning Studio provide a code first environment where analysts write Python or R and they are not a low or no code option. Notebooks require programming knowledge so they do not meet the requirement for business analysts who must avoid coding.

Azure CLI v2 is a command line tool that requires scripting and command knowledge and it is not a graphical, no code platform for training models.

When a question asks for a low or no code solution look for services that advertise guided or automated machine learning rather than notebooks or CLI tools.

The correct option is Compare outputs of the new model with outputs of the existing model.

Compare outputs of the new model with outputs of the existing model is the primary action because it directly verifies that the retrained model does not introduce prediction regressions. Running the same inputs through both models and comparing predictions and evaluation metrics on identical datasets lets you detect degraded accuracy drift or changed behavior before any replacement occurs.

You can perform these comparisons as offline holdout evaluations, as shadow tests where the candidate model runs alongside production traffic without serving responses, or as canary deployments that route a small percentage of traffic to the new model and compare live metrics. These steps reduce risk and provide evidence that the updated model maintains or improves performance.

Automate data preprocessing with a reproducible pipeline is important for consistency and reproducibility but it does not by itself confirm that model predictions will remain as good or better than the existing model. Preprocessing alone will not reveal prediction regressions.

Enable Vertex AI continuous evaluation and model monitoring is useful for detecting issues after a model is deployed to production. It is not the primary predeployment action to ensure a replacement will not reduce performance because monitoring detects problems postdeployment rather than preventing regressions before replacement.

Replace the existing model whenever a predefined replacement rule is met is risky if you do not first validate the candidate by comparing outputs or running safe rollout tests. Automated replacement without direct comparison or staged testing can unintentionally reduce production performance.

Before replacing a production model run a shadow test or offline comparison using the same inputs and metrics as production to catch regressions early.

Google Kubernetes Engine with autoscaling is the correct choice for very low latency and high throughput online model serving.

Google Kubernetes Engine with autoscaling lets you run long running containerized model servers with precise resource requests and limits and it supports Horizontal Pod Autoscaling and the Cluster Autoscaler so you can scale pods and nodes automatically to meet demand while keeping latency low.

Google Kubernetes Engine with autoscaling also provides fine control over load balancing health checks and rolling updates so you can tune deployment topology and reduce tail latency for production traffic.

Cloud Functions is not ideal because serverless functions can experience cold starts and have execution environment limits that make them less suited for very low latency sustained high throughput model serving.

Compute Engine with GPUs can deliver low latency but it requires more manual management of instances and autoscaling and GPUs add cost and operational complexity that are often unnecessary for CPU based inference or when you need rapid horizontal scaling.

Cloud Run provides simple container deployment and autoscaling but it gives you less low level control over scaling behavior and networking than GKE and it can be subject to cold starts and concurrency or scaling characteristics that make it less optimal for the absolute lowest latency at extreme sustained throughput.

When a question requires very low latency and high throughput prefer platforms that support fine grained autoscaling and long running containers such as GKE with autoscaling.

The correct option is score.

The score method on scikit-learn classifiers returns the mean accuracy on the provided test features and labels. You call it as clf.score(X_test, y_test) and it internally generates predictions and compares them to the true labels to compute accuracy, so it is the direct way to obtain the classifier accuracy on a held out test set.

mlflow.log_metric is incorrect because it is an MLflow function for recording metrics to a tracking server and it does not compute a model’s accuracy by itself.

predict is incorrect because predict only returns the model predictions for given inputs and does not compute the accuracy metric. You would need to compare the predictions to the true labels with an accuracy function if you used predict.

fit is incorrect because fit trains the model on data and does not evaluate performance on a test set.

When the question asks for a scikit-learn classifier accuracy think of using estimator.score(X_test, y_test) as it directly returns mean accuracy after training.

Responsible AI Dashboard is the correct option for this scenario because it provides the interactive counterfactual what if analysis needed to find minimal changes that flip a churn prediction to a retained outcome.

The Responsible AI Dashboard includes counterfactual explainers that generate small, realistic perturbations to input features and show which changes would change a model prediction. It integrates with Azure Machine Learning explainers and provides interactive visualizations and metrics so Lena can explore alternative scenarios and refine retention actions based on feasible and measurable feature edits.

The MLflow option is incorrect because MLflow focuses on experiment tracking, model versioning, and deployment management and it does not provide built in interactive counterfactual explainers for what if analysis.

The Azure Machine Learning compute option is incorrect because compute targets supply the processing power for training and scoring and they do not provide an explanation or exploration interface for counterfactual analysis.

The ML Pipelines option is incorrect because pipelines orchestrate data and model workflows and they are not a visualization or explanation tool that generates counterfactual what if scenarios.

When a question asks for counterfactual or what if analysis look for tools that provide explainability dashboards and interactive explainers rather than compute resources or orchestration services.

Exclude specific model families from the candidate pool and Block selected feature transformations and encodings are correct.

With Exclude specific model families from the candidate pool you can narrow the set of algorithms AutoML considers so you can enforce constraints such as interpretability requirements or latency budgets and guide the search toward acceptable model types.

Using Block selected feature transformations and encodings lets you prevent AutoML from applying particular preprocessing steps so you retain control over feature engineering and ensure only allowed transformations and encodings are used.

Override the automatic choice of the primary evaluation metric is incorrect because the primary metric is explicitly configured for the AutoML experiment rather than being something you override mid run, so metric selection is handled at configuration time not by overriding an internal automatic choice.

Disable scaling and normalization of numeric inputs is incorrect because AutoML does not offer a single global toggle to simply turn off numeric scaling. Instead you manage preprocessing by blocking specific transformations and encodings as noted in the correct option.

When a question asks which AutoML controls are available look for answers about restricting model families or blocking transformations. Pay attention to wording that implies explicit configuration like blocking rather than vague global toggles.

The correct answer is Azure AutoML.

Azure AutoML is designed to run automated experiments for tasks such as classification and it performs model selection hyperparameter tuning preprocessing and ensembling without requiring you to write code. You can run it from the Azure Machine Learning studio using a no code experience and it will search many algorithms and configurations to find the highest performing model.

VS Code extensions for Azure ML are focused on a code first workflow inside an integrated development environment and they expect you to write and manage code so they do not meet the requirement to avoid coding.

Azure Machine Learning Designer pipelines provide a drag and drop no code environment for building pipelines but they require you to choose components and configure steps manually and they do not perform the same level of automated model search and tuning that Azure AutoML offers for finding the single best model.

Azure CLI for Machine Learning is a command line tool for scripting and automation and it requires command usage or scripts so it is not a no code GUI solution and therefore it does not satisfy the requirement.

When a question stresses obtaining the highest performing model without writing any code look for an automated training service and emphasize the phrase without writing any code to rule out IDE and CLI options.

The correct answer is az ml compute create -f cluster-config.yml.

This command is the Azure Machine Learning CLI command that provisions a compute target from a YAML definition. Using the az ml compute create -f cluster-config.yml pattern reads the reusable configuration and creates the AmlCompute cluster or other compute target declared in the file so it is the reliable and repeatable approach the team needs.

az ml resource create -f cluster-config.yml is incorrect because the resource subcommand is not the intended command for creating compute clusters and it does not map to the compute create workflow used for AML compute targets.

az ml compute create –name analyticsCluster –sku Standard_D3s_v2 –min-instances 2 –max-instances 6 –type AmlCompute –resource-group analyticsRg –workspace-name analyticsWorkspace is incorrect for this question because it uses explicit flags to create the cluster rather than the reusable YAML file. The command would create a cluster but it does not demonstrate the YAML driven provisioning the documentation needs.

az ml compute add -f cluster-config.yml is incorrect because there is no supported compute add subcommand in the Azure ML CLI for provisioning compute targets. The correct action to add or create compute resources is az ml compute create -f cluster-config.yml.

When a question mentions a reusable YAML file look for a command that accepts the -f or –file option and that explicitly targets the resource type you need, in this case compute.

The correct option is %pip.

%pip is the IPython notebook magic that installs Python packages into the interpreter used by the active notebook kernel. Using the magic ensures the installation targets the kernel that is running the notebook so the newly installed libraries are immediately available to cells without changing other kernels or shared environments.

%conda is the conda equivalent and it installs packages into conda environments when conda is available. It is not the best choice here because the question asks for the magic that reliably installs into the active notebook kernel and %pip is the correct, general answer in Azure Machine Learning notebooks.

%load does not install packages. It loads code from a file or URL into a cell so it is unrelated to package installation and is therefore incorrect.

!pip runs the pip command in the shell and it may invoke a different Python interpreter than the one the notebook kernel is using. Because it can target a different environment it is not the reliable way to install packages into the active kernel.

When you need packages available only to a notebook use %pip inside the notebook so the install targets the active kernel. Avoid relying on !pip when you need certainty about which Python is being modified.