View Full Microsoft AI-900 Exam Dumps and Practice Test Dumps.

 

Question 81

A company has historical examples of customer transactions labeled as fraudulent or legitimate. It wants to train a model to classify future transactions. Which learning approach should be used?

1. Unsupervised learning
2. Reinforcement learning
3. Supervised learning
4. Clustering

Correct Answer: 3

Explanation

Supervised learning is appropriate because the historical transaction data contains known labels indicating whether each transaction was fraudulent or legitimate. The model can learn relationships between the transaction features and these known outcomes. After training, it can use those learned patterns to classify new transactions. This scenario is specifically a classification problem because the expected output belongs to predefined categories. Unsupervised learning does not require labeled outcomes, reinforcement learning learns through rewards and actions, and clustering discovers groups based on similarities without predefined labels.

Question 82

A retailer wants to discover groups of customers based on purchasing patterns, but no customer segments have been defined beforehand. Which technique is most suitable?

1. Regression
2. Clustering
3. Classification
4. Speech recognition

Correct Answer: 2

Explanation

Clustering is an unsupervised learning technique that identifies groups of similar data points without requiring predefined categories. The retailer can provide customer information such as purchase frequency, spending behavior, and product preferences, allowing the algorithm to discover naturally occurring customer segments. Classification would require predefined categories to predict, while regression predicts numerical values. Speech recognition processes spoken language and is unrelated to customer segmentation. Because the retailer wants the AI system to discover the groups rather than predict known groups, clustering is the appropriate technique.

Question 83

A hospital wants to predict whether a patient should be placed into a “high risk,” “medium risk,” or “low risk” category. Which machine learning task is this?

1. Classification
2. Regression
3. Forecasting
4. Clustering

Correct Answer: 1

Explanation

Classification is used when a model assigns an observation to one of several predefined categories. In this scenario, the possible outcomes are high risk, medium risk, and low risk, so the model needs to select a category based on patient information. This differs from regression, which predicts a numerical value, and clustering, which discovers groups without predefined labels. Forecasting generally involves predicting future values over time. Since the hospital has defined the possible risk categories in advance, classification is the appropriate machine learning task.

Question 84

A business wants to estimate the monthly electricity bill for each customer based on historical usage, household size, and other relevant information. What type of machine learning task should it use?

1. Clustering
2. Classification
3. Object detection
4. Regression

Correct Answer: 4

Explanation

Regression is used when a machine learning model needs to predict a numerical value. A customer’s expected monthly electricity bill is a numerical amount, so regression can learn relationships between historical usage, household characteristics, and previous bills. Classification would produce a category, while clustering would identify groups without predefined labels. Object detection is a computer vision capability used to locate objects within images. Because the desired output is a continuous numerical estimate of the customer’s bill, regression is the appropriate machine learning task.

Question 85

In a machine learning project, a customer’s age, annual income, and number of previous purchases are provided to the model as inputs. What are these values called?

1. Labels
2. Features
3. Predictions
4. Metrics

Correct Answer: 2

Explanation

Features are input variables that provide information a machine learning model can use to learn patterns and make predictions. In this example, age, annual income, and previous purchases are characteristics describing each customer, so they can serve as features. A label represents the known target outcome in supervised learning, while a prediction is an output generated by the trained model. Metrics are measurements used to evaluate model performance. Identifying appropriate features is an important part of preparing data for a machine learning solution.

Question 86

A dataset contains information about houses along with their actual selling prices. If the model is being trained to predict selling prices, what does the selling price represent?

1. Feature
2. Label
3. Prediction
4. Hyperparameter

Correct Answer: 2

Explanation

The actual selling price is the label because it represents the known target value that the supervised learning model is being trained to predict. Other information, such as house size, number of bedrooms, and location, could be used as features. During training, the model learns relationships between those input features and the known selling prices. After training, it can use features from a new house to generate a predicted selling price. Therefore, the actual historical selling price serves as the label in this regression problem.

Question 87

Why should a machine learning model be evaluated using data that was not used to train it?

1. To determine how well it generalizes to unseen examples
2. To make the training data larger
3. To remove all model features
4. To guarantee perfect predictions

Correct Answer: 1

Explanation

Evaluating a model on unseen data helps determine whether it has learned patterns that generalize beyond the examples used during training. A model may perform very well on training data while failing to make accurate predictions on new examples. Using separate evaluation data provides a more realistic measurement of performance. It does not guarantee perfect predictions because all models have limitations and may encounter unfamiliar conditions. The purpose is to understand how the model is likely to behave when it receives data that was not part of its training process.

Question 88

A model memorizes specific patterns in its training data and performs poorly when given new examples. What problem is this an example of?

1. Underfitting
2. Overfitting
3. Classification
4. Feature engineering

Correct Answer: 2

Explanation

Overfitting occurs when a machine learning model learns the training data too closely and fails to generalize effectively to new examples. The model may capture noise or details that are specific to the training dataset rather than learning broader patterns. As a result, training performance can be very high while performance on unseen data is significantly worse. Developers can investigate approaches such as using more representative data, reducing unnecessary complexity, or applying suitable regularization techniques. Underfitting has the opposite general pattern, where the model fails to learn enough from the training data.

Question 89

A model is so simple that it cannot capture important relationships in its training data and performs poorly on both training and test datasets. Which issue is most likely present?

1. Overfitting
2. Data encryption
3. Underfitting
4. Object detection

Correct Answer: 3

Explanation

Underfitting occurs when a model is not capable of learning enough of the relevant patterns in the data. This can happen when the model is too simple, the selected features are insufficient, or the training process does not allow the model to learn adequately. An underfitted model typically performs poorly on both training and unseen data. Overfitting generally produces strong training performance but weaker performance on new data. Therefore, consistently poor performance caused by an overly simple model is a typical indication of underfitting.

Question 90

A data science team needs a managed Azure environment for training, deploying, and monitoring machine learning models. Which service should it use?

1. Azure Machine Learning
2. Azure AI Speech
3. Azure AI Vision
4. Azure DNS

Correct Answer: 1

Explanation

Azure Machine Learning provides a managed environment for developing, training, deploying, and managing machine learning solutions. It can support activities across the machine learning lifecycle, including data preparation, model training, evaluation, deployment, and monitoring. Azure AI Speech focuses on speech-related capabilities, Azure AI Vision handles computer vision workloads, and Azure DNS provides domain name resolution. Therefore, Azure Machine Learning is the most appropriate service when an organization needs a broader platform for managing machine learning models and workflows.

Question 91

A data scientist wants Azure to automatically try different machine learning algorithms and settings to identify a suitable model for a particular dataset. Which capability can help with this?

1. OCR
2. Automated machine learning
3. Speech synthesis
4. Image captioning

Correct Answer: 2

Explanation

Automated machine learning, commonly called AutoML, can automate parts of the machine learning model development process. Depending on the selected task and configuration, it can evaluate different algorithms, preprocessing approaches, and model settings to identify candidates that perform well against specified evaluation criteria. This can reduce the amount of manual experimentation required during model development. OCR extracts text from images, speech synthesis generates spoken audio, and image captioning creates descriptions of visual content. AutoML is therefore the capability most relevant to automated model experimentation and selection.

Question 92

An organization wants to organize several machine learning steps, including data preparation, training, evaluation, and deployment, into a repeatable workflow. What can help accomplish this?

1. A machine learning pipeline
2. Sentiment analysis
3. OCR
4. Speech recognition

Correct Answer: 1

Explanation

A machine learning pipeline organizes multiple steps of an AI or machine learning workflow into a structured and repeatable process. A pipeline can include activities such as preparing data, training a model, evaluating its performance, and performing other processing tasks. This can improve consistency and make workflows easier to automate and maintain. Sentiment analysis analyzes the emotional tone of text, OCR extracts text from images, and speech recognition converts spoken language into text. Because the requirement involves connecting several machine learning stages into a repeatable process, a pipeline is appropriate.

Question 93

An online shopping application needs to return a product recommendation immediately after a customer clicks an item. Which inference pattern is most appropriate?

1. Batch inference
2. Online inference
3. Offline data preparation
4. Historical reporting

Correct Answer: 2

Explanation

Online inference is designed for applications that require predictions or results in response to individual requests, often with low latency. In this scenario, the shopping application needs to generate a recommendation immediately after the customer interacts with a product. An online endpoint can process the request and return a prediction without waiting for a large collection of requests to be processed together. Batch inference is more suitable when many records can be processed as a group. Therefore, online inference is appropriate for this interactive recommendation scenario.

Question 94

A company wants to score two million customer records every night and does not require an immediate response for each individual record. Which inference pattern is suitable?

1. Online inference
2. Real-time speech processing
3. Batch inference
4. Image captioning

Correct Answer: 3

Explanation

Batch inference is designed for processing large collections of data together rather than responding immediately to individual requests. In this scenario, the company can process the two million customer records overnight because an immediate response is not required. Batch processing can be useful for scheduled scoring, reporting, large-scale data analysis, and other workloads where results can be generated periodically. Online inference is more appropriate for low-latency individual requests. Image captioning and speech processing address different AI workloads and are unrelated to the described large-scale scheduled scoring requirement.

Question 95

A classification model correctly predicts 92 out of 100 customer categories. Which metric could directly express this overall proportion of correct predictions?

1. Accuracy
2. Recall
3. Precision
4. Mean absolute error

Correct Answer: 1

Explanation

Accuracy represents the proportion of predictions that a classification model gets correct out of all evaluated predictions. If a model correctly predicts 92 out of 100 examples, its accuracy is 92 percent. Accuracy provides an overall measure of correct predictions but may not fully describe model behavior when classes are highly imbalanced. Precision focuses on how many predicted positive cases are actually positive, while recall measures how many actual positive cases were identified. Mean absolute error is generally associated with numerical prediction errors in regression problems.

Question 96

A fraud detection model identifies 90 percent of the fraudulent transactions that actually occurred. Which metric most directly describes this result?

1. Precision
2. Accuracy
3. Recall
4. Mean absolute error

Correct Answer: 3

Explanation

Recall measures the proportion of actual positive cases that a classification model successfully identifies. If a fraud detection model finds 90 percent of all fraudulent transactions that occurred, its recall is 90 percent. Recall is particularly important when missing positive cases can have significant consequences. Precision, in contrast, focuses on how many predicted positive cases are actually positive. Accuracy measures overall correctness, while mean absolute error is commonly used for regression. Therefore, recall is the metric that most directly describes the stated fraud detection result.

Question 97

A security model flags 100 transactions as fraudulent, but only 80 of those transactions are actually fraudulent. Which metric measures this proportion of correct positive predictions?

1. Recall
2. Precision
3. Accuracy
4. R-squared

Correct Answer: 2

Explanation

Precision measures the proportion of predicted positive cases that are actually positive. In this scenario, the model identifies 100 transactions as fraudulent and 80 are truly fraudulent, giving a precision of 80 percent. Precision is especially useful when false positive predictions can create significant costs or unnecessary investigations. Recall instead measures the proportion of all actual fraudulent transactions that the model successfully identifies. Accuracy considers all predictions, while R-squared is associated with regression performance. Therefore, precision is the appropriate metric for this situation.

Question 98

A regression model predicts house prices, and the organization wants to measure the average absolute difference between predicted and actual prices. Which metric should it consider?

1. Recall
2. Accuracy
3. Precision
4. Mean absolute error

Correct Answer: 4

Explanation

Mean absolute error, or MAE, measures the average absolute difference between predicted values and actual values. For a house-price regression model, it can indicate how far predictions typically differ from the actual selling prices. A lower MAE generally means that predictions are closer to the actual values, although the suitability of a metric depends on the business problem and evaluation requirements. Precision, recall, and accuracy are commonly associated with classification problems rather than measuring continuous numerical prediction errors.

Question 99

A business wants to extract fields such as customer names, invoice numbers, dates, and totals from invoices automatically. Which Azure capability is designed for this type of document processing?

1. Azure AI Document Intelligence
2. Azure AI Speech
3. Azure AI Vision only
4. Azure DNS

Correct Answer: 1

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract structured information from documents such as invoices, receipts, forms, and other business records. It can use prebuilt models for supported document types and can also support custom document processing scenarios. This allows extracted information to be passed into business applications and workflows. Azure AI Speech focuses on spoken language, Azure AI Vision provides broader computer vision capabilities, and Azure DNS handles domain name resolution. Document Intelligence is therefore the appropriate capability for structured invoice information extraction.

Question 100

A generative AI application retrieves relevant information from an organization’s documents and provides that information to the model before generating an answer. Which architecture does this describe?

1. Object detection
2. Retrieval-augmented generation
3. Image classification
4. Reinforcement learning

Correct Answer: 2

Explanation

Retrieval-augmented generation, or RAG, combines information retrieval with generative AI. When a user asks a question, the application retrieves relevant information from a connected source such as organizational documents, a knowledge base, or another approved data repository. The retrieved information is then provided as context to the generative model, which uses that context when producing its response. RAG can help improve relevance and provide a way to ground responses in external information. Object detection, image classification, and reinforcement learning describe different AI workloads or approaches.