ML Techniques and Real-World Use Cases

Start with the shape of the answer

Most "which technique?" questions are decided before any algorithm is named, by looking at what the answer has to look like. A number, a known label, or a group nobody has named yet — those are three different shapes, and each one belongs to a different family.

Start with the numeric shape. Regression predicts a numeric value, such as a house price. Choose regression when the target is a quantity on a numeric scale rather than a named category.

A demand estimate, temperature estimate, or predicted delivery time has this shape. Notice what they share: the answer could be 12, or 12.4, or 340, and every value in between is meaningful.

Now the categorical shape. Classification predicts a category, such as spam or not spam. Choose classification when the allowed output is a class label.

Image classification, sentiment categories, and a potentially fraudulent transaction label are classification-shaped problems. The set of possible answers is written down in advance, and the model's job is to pick from that set.

Discover groups that nobody labeled

The third shape is the one people forget, because it does not predict anything at all. Clustering groups similar observations without requiring predefined class labels. Choose clustering when the task is to discover natural segments, such as groups of customers with similar behavior, rather than predict a known number or label.

The absence of labels is the giveaway. If a retailer already had a list of segments and wanted new customers sorted into them, that would be classification. When the segments themselves are the deliverable, it is clustering.

Put the three together and the rule is short. Select regression for a number, classification for a predefined label, and clustering for discovered groups.

The business wording is decisive: estimate how much, decide which known class, or discover which observations belong together. Those three verbs — estimate, decide, discover — usually appear verbatim in the requirement, and they are more reliable than any hint about the data or the tooling.

Applications that see and read

Above the technique families sit the application types, and those are chosen by the kind of data going in. Start with pictures.

Computer vision extracts information from images or video and supports tasks such as object detection, image classification, and visual anomaly detection. Visual inspection of manufactured products is a computer vision application: a camera on the line, and a model that reports what is present, what is missing, or what looks wrong.

Vision is not one task but a family of them. Object detection says what is in the frame and where; image classification labels the whole image; visual anomaly detection reports that something departs from the normal appearance without necessarily naming it.

Then text. Natural language processing helps computers understand human language for tasks such as translation, intelligent search, and text analysis. Extracting entities, key phrases, topics, or sentiment from documents is an NLP application.

That last list is worth memorizing as a group, because it is what "analyze this text" usually means in practice:

• Entities — the people, places, products and dates a document mentions.
• Key phrases — the expressions that carry the document's substance.
• Topics — the themes that let a pile of files be organized.
• Sentiment — whether the writer is positive or negative.

Applications that listen

Speech recognition converts spoken audio into text. Transcribing a recorded call or producing a real-time transcript is a speech-recognition application; generating audio from text is a different direction.

Direction is the whole point here, and it is the detail scenarios are built on. Audio arrives, text comes out. The reverse trip — text arrives, audio comes out — is speech synthesis, a separate capability that happens to sit next to it in every service catalog.

Both a recorded file and a live stream count as audio input, so "real time" changes the latency requirement, not the family of the application.

Applications that personalize and flag

Recommendation systems use activity and item information to produce personalized product or content suggestions. Two inputs are needed, not one: what this user has done, and what the catalog contains.

The required output is a ranked or tailored set of items for a user, not a demand forecast for the whole business. That contrast is the trap in most scenarios: both a recommender and a forecaster look at historical behavior, but one answers "what should this person see next" and the other answers "how much will everyone buy".

Fraud detection identifies potentially fraudulent activity so that suspicious events can be caught or reviewed. Its output commonly supports a fraud label or risk score for an event.

The unit of the answer is the event: this payment, this login, this claim. A fraud application does not personalize anything and does not project a total; it scores what just happened so that it can be blocked, queued or waved through.

Applications that look ahead and look up

Forecasting predicts future business outcomes from time-series data and relevant variables. Product demand, resource needs, and financial performance are forecasting examples.

Two ingredients define it: a series of values recorded over time, and a horizon in the future. If the requirement mentions "next quarter", "next year" or "the coming season" together with a quantity, forecasting is almost always the answer.

Knowledge-base applications retrieve relevant information from an organized body of content to answer a user's information need.

The result already exists somewhere; the work is finding and returning it. This differs from forecasting a future value and from discovering unlabeled customer segments — nothing is being estimated and nothing is being grouped.

Coordinate the whole task

Agentic AI manages intent, completes multi-step work, and can validate progress toward a goal. It is appropriate when the application must orchestrate actions across a task rather than return a single classification, forecast, transcript, or recommendation.

The tell is the plural. One prediction is not agentic, however sophisticated it is; a sequence of actions chosen and checked against an objective is. A tool that reads a request, edits several files, runs the tests and confirms that the goal was met is doing work no single classifier or forecaster does.

Key takeaways

Start with the data type, then identify the required output; similar business domains can still require different applications. Vision consumes visual content, NLP analyzes language, speech recognition transcribes audio, recommendations personalize items, fraud detection flags suspicious activity, forecasting estimates future outcomes, knowledge bases retrieve grounded information, and agentic AI coordinates actions.

• Use regression for numeric targets, classification for predefined labels, and clustering for discovered groups.
• Vision handles visual content, NLP handles language, and speech recognition maps audio to text.
• Recommendations personalize items, fraud detection flags suspicious events, and forecasting predicts future outcomes.
• Knowledge bases retrieve information, while agentic AI coordinates multi-step actions.