ML Techniques

AI-900 Domain 2 ("Describe fundamental principles of machine learning on Azure") expects you to identify the right machine-learning technique from a one-paragraph scenario, and Microsoft organizes the whole space with a single decision: does the training data carry known labels?^[1]

The figure groups the techniques the way Microsoft does: machine learning splits into two families by whether the data is labeled, and each family contains the specific techniques this page then walks one by one.

Supervised learning: the training data already contains the answers

In supervised learning, the training data includes both feature values (the inputs) and known label values (the correct answers). The model is trained "by determining a relationship between the features and labels in past observations, so that unknown labels can be predicted for features in future cases." Practically: every training row has the right answer attached, a price, a yes/no, a species, and the model learns to reproduce it. Supervised learning splits into two forms by what kind of label it predicts:

• Regression: the label is a numeric value.
• Classification: the label is a category (a class). Two sub-shapes: binary (one of two) and multiclass (one of several).

Unsupervised learning: features only, no answers

Unsupervised learning trains on data "that consists only of feature values without any known labels." There is no answer column; the algorithm finds structure or relationships among the observations themselves. The most common unsupervised technique is:

• Clustering: group similar observations into discrete clusters based purely on feature similarity.

The mechanical test for the exam

When a stem describes the data, ask: is there a known answer attached to each training row?

If the scenario…	…the technique is	Learning type
has labeled rows and the answer is a number	Regression	Supervised
has labeled rows and the answer is one of two classes	Binary classification	Supervised
has labeled rows and the answer is one of several classes	Multiclass classification	Supervised
has no labels and you want to discover groupings	Clustering	Unsupervised

This single table answers the large majority of "which technique" questions on AI-900. The rest of this page makes each row concrete with Microsoft's own canonical examples so you recognize them verbatim if the exam quotes them.

Regression^[2] is "a form of supervised machine learning in which the label predicted by the model is a numeric value." You train it on past observations that pair features with a known number, and it predicts a continuous value for new inputs.

Microsoft's canonical AI-900 examples (memorize these)

• Ice cream sales: predict the number of ice creams sold on a given day, based on temperature, rainfall, and windspeed.
• Property price: predict the selling price of a property based on its size in square feet, number of bedrooms, and socio-economic metrics for its location.
• Fuel efficiency: predict a car's fuel efficiency (miles-per-gallon) based on engine size, weight, width, height, and length.

Every one of these asks for an amount: a count, a price, a rate. That is the regression signature.

The exam tell

The word that gives regression away answers "how many / how much / what value / what amount." Any scenario asking to forecast a quantity, amount, price, count, or score points to regression, not classification. "How many units will we sell next month?", "What will this house sell for?", "What temperature will the reactor reach?" are all regression.

The classic trap

Don't let a number that names a category fool you. "Predict whether a customer is in tier 1, 2, or 3" is classification even though the labels look numeric. The values are discrete classes, not a continuous quantity. Regression is for a value on a continuous scale; if the output is a label drawn from a fixed set of categories, it is classification.

Classification^[3] is "a form of supervised machine learning in which the label represents a categorization, or class." Like regression it needs labeled training data, but the label is a category, not a number. Microsoft divides it into two scenarios, and the rarer multilabel variant sits under multiclass; the figure shows that breakdown before each is described.

Binary classification: one of two mutually exclusive outcomes

Binary classification predicts "one of two mutually exclusive outcomes," a true/false or positive/negative answer for a single class. Microsoft's examples:

• Whether a patient is at risk for diabetes, based on clinical metrics like weight, age, and blood glucose level.
• Whether a bank customer will default on a loan, based on income, credit history, age, and other factors.
• Whether a mailing-list customer will respond positively to a marketing offer, based on demographics and past purchases.

The shape is always two options: yes/no, will/won't, is/isn't.

Multiclass classification: one label out of several possible classes

Multiclass classification^[4] "extends binary classification to predict a label that represents one of multiple possible classes." Microsoft's examples:

• The species of a penguin (Adelie, Gentoo, or Chinstrap) based on physical measurements.
• The genre of a movie (comedy, horror, romance, adventure, or science fiction) based on cast, director, and budget.

Standard multiclass predicts mutually exclusive labels. "A penguin can't be both a Gentoo and an Adelie," so the model picks exactly one class.

Multilabel: the rarer variant (recognize it, don't overthink it)

Microsoft notes that some algorithms train multilabel models, "in which there may be more than one valid label for a single observation," for example a movie tagged as both science fiction and comedy. If a stem explicitly says an item can carry more than one category at once, that is multilabel; if each item gets exactly one of several categories, it is standard multiclass.

The exam tell

Classification answers "which category / which class / true or false." Count the possible answers: two mutually exclusive outcomes → binary; several mutually exclusive classes → multiclass; multiple simultaneous tags → multilabel. Contrast with regression, which answers "how much."

Clustering^[5] is "the most common form of unsupervised machine learning." A clustering algorithm "identifies similarities between observations based on their features, and groups them into discrete clusters," with no previously known label to learn from. Microsoft's examples:

• Group similar flowers based on their size, number of leaves, and number of petals.
• Identify groups of similar customers based on demographic attributes and purchasing behavior (customer segmentation).

Clustering vs multiclass classification: the single most important distinction here

Both assign observations to discrete groups, which is exactly why AI-900 loves to test the difference. The deciding factor is whether the classes are known in advance:

• Classification is supervised: you already know the classes, and the algorithm learns the mapping from features to those known labels.
• Clustering is unsupervised: there is no predefined cluster label; the algorithm groups observations "based purely on similarity of features."

So a scenario that supplies known class labels in the training data is classification, not clustering. A scenario with no labels where the goal is to discover natural groupings is clustering.

Bonus insight Microsoft calls out

Clustering is often used first, to discover what classes even exist. Microsoft's example: cluster customers into segments, analyze and label those segments ("high value – low volume," "frequent small purchaser"), then use that labeled data to train a classification model that predicts which segment a new customer belongs to. The figure traces that pipeline: cluster to discover → label → classify to assign.

The exam tell

No labels + "segment / group / find natural groupings / discover structure" → clustering. Known labels + "assign to one of these categories" → classification.

Deep learning^[6] is "an advanced form of machine learning that tries to emulate the way the human brain learns" by building an artificial neural network. For AI-900 you don't need the calculus, you need the vocabulary and the high-level mechanism.

The building blocks

• Neuron: a function that operates on an input value (x) and a weight (w).
• Activation function: wraps each neuron and "determines whether to pass the output on" to the next layer.
• Layers: neural networks stack multiple layers of neurons, "essentially defining a deeply nested function." That nesting, depth, is why the technique is called deep learning and the models are called deep neural networks (DNNs).

What it actually does (same job as the other techniques)

Microsoft is explicit: "just like other machine learning techniques… deep learning involves fitting training data to a function that can predict a label (y) based on the value of one or more features (x)." Deep neural networks can do regression and classification, and are the foundation of more specialized models for computer vision and natural language processing.

How a DNN learns, in one breath

The figure traces Microsoft Learn's own deep-learning training loop, a single cycle repeated over many epochs:

1. Feature values (x) are fed forward through the layers; each neuron applies its weight and activation, producing predicted outputs (ŷ).
2. A loss function compares the predicted ŷ to the known y and aggregates the error into a single loss value.
3. An optimizer adjusts the weights, typically via gradient descent, backpropagated through the layers, to reduce that loss.
4. The process repeats over many iterations (epochs) until loss is minimized.

The trained model is simply "the final weight values that result in the most accurate predictions."

The exam framing

Think of deep learning as many layers of weighted neurons trained by minimizing loss. When a stem mentions "neural network," "deep neural network / DNN," "layers," "weights," or "emulating the human brain," the answer is deep learning. Deep learning is how the more advanced Azure AI capabilities (vision, language, generative models) are built, but it is still ordinary supervised learning underneath, predicting a label from features.

AI-900 rarely asks for definitions in the abstract. It gives you a workload and asks which technique fits. Run every stem through this two-step filter.

Step 1: Are the training rows labeled?

• Labels present → supervised → go to Step 2.
• No labels, want to discover groups → clustering (unsupervised). Done.

Step 2: Is the predicted label a number or a category?

• A number (price, count, amount, rate, score) → regression.
• A category → classification: two mutually exclusive options → binary; several mutually exclusive options → multiclass; multiple simultaneous tags → multilabel.

Worked scenario → technique mappings

Scenario stem	Technique	Why
"Predict tomorrow's ice cream sales from weather data"	Regression	Labeled history + numeric output (a count)
"Estimate a house's selling price from its features"	Regression	Labeled history + numeric output (a price)
"Flag whether a transaction is fraudulent (yes/no)"	Binary classification	Two mutually exclusive labeled classes
"Predict whether a customer will default on a loan"	Binary classification	True/false labeled outcome
"Identify a penguin's species: Adelie, Gentoo, or Chinstrap"	Multiclass classification	One of several known labeled classes
"Tag a movie with all applicable genres at once"	Multilabel classification	Multiple valid labels per item
"Segment customers into similar groups (no predefined groups)"	Clustering	No labels; discover groupings by similarity
"Recognize objects in photos using a deep neural network"	Deep learning (classification under the hood)	Neural-network keywords; predicts a class

Highest-yield traps to internalize

• "How much / how many" is regression, "which category" is classification. Don't pick classification for a continuous-number question.
• Numeric-looking class labels are still classification. Tiers 1/2/3 are categories, not a continuous quantity.
• Clustering vs multiclass: both make groups, but only classification has known labels. "No predefined labels" → clustering; "assign to these known classes" → classification.
• Regression and classification require labels. If the stem says the data has no known answers, neither applies. It must be clustering.
• Deep learning is not a separate problem type. It still does regression or classification; the giveaway is neural-network / layers / weights wording, not a new kind of output.