Microsoft AI-900

Correct answer: A (classification). Predicting whether a loan will be repaid is a supervised learning problem where the target is a discrete label, such as “repaid” vs “not repaid” (default). That is the definition of classification: learning from historical labeled examples (past loans with known outcomes) to predict a category for a new case. Why not regression (B): Regression is used when the output is a continuous numeric value, such as predicting the exact loan loss amount, interest rate, or customer lifetime value. While a credit model might output a probability (a number), the core task described—“whether a loan will be repaid”—is a categorical decision. Why not clustering (C): Clustering is unsupervised and groups data points by similarity without known labels (e.g., segmenting customers into risk groups when you don’t have repayment outcomes). Here, the bank explicitly wants a known outcome prediction, which requires labeled training data and classification.

“Correctly predicted positives” means True Positives (TP): cases where the model predicted class 1 and the actual class is also 1. In the confusion matrix, that is the intersection of Predicted = 1 (row) and Actual = 1 (column), which is the top-left cell. The value shown there is 11, so there are 11 correctly predicted positives. Why the other options are wrong: - 5 is False Positives (predicted 1, actual 0). - 1,033 is False Negatives (predicted 0, actual 1). - 13,951 is True Negatives (predicted 0, actual 0).

False Negatives (FN) are cases where the model predicted the negative class (0) but the actual class is positive (1). This represents missed positives, which is often critical in scenarios like fraud detection or medical screening. In the confusion matrix, FN is located at Predicted = 0 (row) and Actual = 1 (column), which is the bottom-left cell. The value in that cell is 1,033, so there are 1,033 false negatives. Why the other options are wrong: - 11 is True Positives (predicted 1, actual 1). - 5 is False Positives (predicted 1, actual 0). - 13,951 is True Negatives (predicted 0, actual 0).

Correct answer: B. features. Features are the input data values (columns) that the model uses to learn patterns and make predictions. For example, in a house-price model, square footage, number of bedrooms, and location are features because they influence the predicted price. Why the others are wrong: - A. dependant variables: In most ML terminology, the dependent variable is the target/output you are trying to predict (often synonymous with the label), not the inputs that influence the prediction. - C. identifiers: Identifiers (like CustomerID) uniquely identify records but typically should not be used as predictive inputs because they don’t represent generalizable patterns; they can also introduce data leakage. - D. labels: A label is the known outcome/target value used for training in supervised learning (for example, “price” or “spam/not spam”), not the input values that influence the prediction.

Forecasting housing prices based on historical data is not anomaly detection. The key phrase is “forecasting … based on historical data,” which indicates a time-series forecasting problem: you use past values (and possibly external variables like interest rates, inventory, seasonality) to predict future prices. The objective is to estimate expected future values, not to identify rare deviations. Anomaly detection would instead focus on finding unusual house prices or transactions that deviate from normal patterns (for example, a sale price far above comparable properties, potentially indicating data errors or fraud). In AI-900, forecasting and anomaly detection are distinct workload types: forecasting predicts what will happen; anomaly detection flags what is unusual compared to normal behavior. Therefore, the correct answer is No.

Identifying suspicious sign-ins by looking for deviations from usual patterns is a classic anomaly detection scenario. The goal is to detect rare or unusual events—such as logins from atypical locations, impossible travel (two distant sign-ins in a short time), unusual device fingerprints, abnormal login times, or spikes in failed attempts—relative to a baseline of normal user behavior. This aligns directly with the definition of anomaly detection: finding observations that differ significantly from the majority. It is often used in cybersecurity and fraud detection because anomalies can indicate compromised accounts or malicious activity. This is not primarily forecasting (predicting a future numeric value) or standard classification unless you already have labeled examples of “suspicious” vs. “normal.” Even when classification is used, the underlying concept in the prompt—“deviations from usual patterns”—maps to anomaly detection. Therefore, the correct answer is Yes.

Predicting whether a patient will develop diabetes based on the patient’s medical history is not anomaly detection; it is classification. The output is a discrete label (for example, “will develop diabetes: yes/no”), and the model is trained on historical patient records where the outcome is known. This is a supervised learning task. Anomaly detection would be more appropriate if the goal were to identify unusual patient measurements or rare patterns that deviate from a normal population (for example, detecting abnormal lab results that might indicate an undiagnosed condition). But the prompt is explicitly about predicting a specific outcome (diabetes development), which is a typical binary classification use case. In AI-900, remember: classification predicts categories, regression predicts numbers, forecasting predicts future time-based values, and anomaly detection flags rare deviations. Since the scenario is outcome prediction (yes/no), the correct answer is No.

Sentiment analysis is the correct feature because the scenario is explicitly about understanding how upset a customer is, which is an opinion/emotion classification problem. Azure AI Language sentiment analysis evaluates text and returns sentiment labels (such as negative/neutral/positive) and often confidence scores; it can also provide sentence-level sentiment to pinpoint the most negative parts of a ticket. Why the others are wrong: - Entity recognition extracts structured items (like dates, names, products) but does not measure emotion. - Key phrase extraction identifies important terms/topics (e.g., “login failure”, “billing issue”) but not how the customer feels. - Language detection only identifies the language of the text and does not assess sentiment. In a ticketing system, sentiment is commonly used for prioritization/escalation rules (e.g., highly negative tickets routed to senior support).

Key phrase extraction is the best match for summarizing important information from a support ticket. This feature pulls out the most relevant words and phrases that represent the main topics of the text (for example: “VPN connection”, “authentication error”, “password reset”). That effectively creates a lightweight “summary” suitable for tagging, search indexing, routing, and analytics dashboards. Why the others are wrong: - Sentiment analysis summarizes emotional tone, not the key technical content. - Entity recognition extracts specific entity types (dates, organizations, products), but it may miss the broader topic phrases and is not intended as a general summary. - Language detection only identifies the language and provides no content summary. On AI-900, “summarize important information” in the context of Text Analytics typically maps to key phrase extraction rather than abstractive summarization (which is a different capability).

Entity recognition is correct for extracting key dates from the support ticket because dates are a standard entity type that Named Entity Recognition (NER) can detect and return in structured form. For example, it can identify values like “January 15, 2026”, “last Friday”, or “10/12/2025” as date/time entities, enabling you to populate ticket fields, correlate incidents, or validate SLAs. Why the others are wrong: - Key phrase extraction might surface a date-like phrase, but it is not designed to reliably classify and structure dates as entities. - Sentiment analysis evaluates tone and does not extract factual items like dates. - Language detection only determines the language of the ticket. In practice, entity recognition is used to extract many structured fields (dates, product names, locations, people) from unstructured ticket text to support automation and reporting.

Correct answer: C (reliability and safety). “Handling of unusual or missing values” is fundamentally about how robust and dependable an AI system is when it encounters imperfect or unexpected input. Reliability and safety in Microsoft’s Responsible AI principles focuses on ensuring AI systems perform consistently under normal conditions and degrade gracefully under abnormal conditions (for example, null values, outliers, corrupted records, or unexpected formats). It also includes preventing harmful outcomes when the system is uncertain or failing. Why the others are wrong: - A (inclusiveness) is about designing systems that empower and include people of diverse abilities and backgrounds (accessibility, equitable user experience), not about input anomaly handling. - B (privacy and security) concerns protecting sensitive data, access control, encryption, and resisting data leakage or adversarial attacks. - D (transparency) is about explainability and communicating how the system works, its limitations, and when it may be unreliable—related, but not the core principle for handling missing/unusual values. In practice, reliability/safety is addressed via validation, fallback logic, monitoring, and testing with edge cases.

Correct: E (Reliability and safety). The description emphasizes that AI systems should behave consistently with their intended design, handle unexpected conditions, and resist harmful manipulation. These are classic reliability/safety concerns: robustness, fault tolerance, resilience to adversarial inputs, and safe operation under edge cases. Why others are wrong: - A (Accountability) is about who is responsible and ensuring oversight and governance, not the technical/operational robustness of the model. - B (Fairness) focuses on avoiding discriminatory outcomes across groups. - C (Inclusiveness) focuses on accessibility and ensuring solutions work for people with diverse abilities and needs. - D (Privacy and security) focuses on protecting data and controlling data usage, not primarily on model robustness and safe behavior.

Correct: A (Accountability). The key phrase is that AI decisions “can be overridden by humans.” This is human-in-the-loop governance: ensuring there is appropriate oversight, escalation paths, and the ability for people to intervene when an automated decision is incorrect or inappropriate. Accountability also implies clear ownership for outcomes and processes for auditing and remediation. Why others are wrong: - E (Reliability and safety) is about the system operating safely and robustly, but the explicit requirement for human override maps more directly to accountability and governance. - D (Privacy and security) concerns data protection and consent, not decision override. - B (Fairness) concerns bias and equitable outcomes. - C (Inclusiveness) concerns accessibility and designing for diverse users.

Correct: D (Privacy and security). The description is explicitly about giving consumers information and controls over the collection, use, and storage of their data. That aligns directly with privacy principles (transparency, consent, data minimization, appropriate retention) and security principles (protecting data at rest/in transit, access control). Why others are wrong: - A (Accountability) is about oversight and responsibility for AI outcomes, not user controls over personal data. - E (Reliability and safety) is about robust and safe system behavior. - B (Fairness) is about avoiding biased outcomes. - C (Inclusiveness) is about ensuring the solution is usable and beneficial for people with diverse needs, not data governance.

Correct answer: A. Classifying email messages as work-related or personal is a classic NLP task called text classification. The input is unstructured text (email subject/body), and the model or service determines a label/category based on language features (words, phrases, semantics). In Azure, this aligns with NLP capabilities such as Azure AI Language (text classification) or custom text classification. Why the others are wrong: B (predict the number of future car rentals) is forecasting/regression on historical numeric data (time series), which is a machine learning workload, not NLP. C (predict which website visitors will make a transaction) is typically a binary classification ML problem using clickstream/session attributes and customer features; it doesn’t require language understanding. D (stop a process in a factory when extremely high temperatures are registered) is a sensor/IoT threshold or anomaly detection scenario; it involves telemetry and operational rules rather than processing human language.

Correct answer: B. Azure Container Instances. In Azure Machine Learning Designer, when you want to deploy a real-time inference pipeline so others can call it (typically via a REST endpoint), you deploy it to a compute target that hosts the web service. Azure Container Instances (ACI) is the standard, simplest managed option for real-time endpoints—commonly used for development, testing, and low-to-moderate traffic production scenarios. It requires minimal infrastructure management and is frequently referenced in AI-900-level content as the default real-time deployment target. Why the others are wrong: A. A local web service is for local testing/debugging and is not a cloud-hosted service intended for broad consumption. C. AKS is also a valid real-time hosting option, but it’s typically chosen for production at scale (high availability, autoscaling, advanced networking). The question doesn’t indicate those requirements, so ACI is the most appropriate. D. Azure ML compute is primarily for training and batch processing; it isn’t the typical hosting target for a deployed real-time web service endpoint.

Correct answer: B. Custom Vision. Azure AI Custom Vision is designed specifically to train custom computer vision models using your own images. For object detection, you upload images, draw bounding boxes around objects, label them, and train a model that can locate and classify objects within new images. This directly matches the requirement “train an object detection model by using your own images.” Why the others are wrong: - A. Computer Vision (Azure AI Vision) provides prebuilt models (for example, tagging, OCR, and general detection) but is not the primary service for training a custom object detector with your own labeled dataset in the AI-900 context. - C. Form Recognizer (Azure AI Document Intelligence) is for extracting structured data from documents (forms, invoices, receipts) using OCR + layout/field extraction, not general object detection in images. - D. Video Indexer is for extracting insights from video/audio (transcripts, faces, keywords, scenes) rather than training an object detection model from your own still images.

Yes. Labeling is the process of associating each training example with the correct known value (the “ground truth”). In supervised learning, these labels are what the model learns to predict—for example, tagging images as “cat” or “dog,” marking emails as “spam” or “not spam,” or assigning a numeric target value for regression such as “house price.” Labeling can be done manually by humans, programmatically (when rules or existing systems provide ground truth), or via assisted labeling tools. Without labels, you typically move into unsupervised learning (clustering, dimensionality reduction) or self-supervised approaches, which are different problem types. Therefore, the statement correctly describes labeling as tagging training data with known values.

No. You should not evaluate a model using the same data used to train it because it does not measure generalization. Training performance often looks better than real-world performance, especially if the model overfits (learns noise or memorizes examples). Proper evaluation uses a separate test dataset (or validation dataset) that the model has not seen during training. Common approaches include train/test splits, train/validation/test splits, and cross-validation. This helps estimate how the model will perform on new data and supports better decisions about model selection and hyperparameter tuning. Using training data for evaluation can lead to deploying a model that appears accurate in development but fails in production.

No. Accuracy is not always the primary metric; the best metric depends on the task and the consequences of different error types. In imbalanced classification (e.g., fraud detection), a model can achieve high accuracy by always predicting the majority class while missing the minority class entirely. In such cases, recall (sensitivity), precision, F1-score, and PR-AUC are often more informative. If false positives are costly (flagging legitimate transactions as fraud), precision may matter more; if false negatives are costly (missing fraud), recall may be prioritized. For regression problems, accuracy is not the standard metric at all—metrics like MAE or RMSE are used. Therefore, accuracy is not universally the primary measure of model performance.