Inputs, targets, and predictions

By Segmentation Fault Academy · Published 3 October 2026

Inputs are the information a model uses. In supervised learning, the target is the outcome you want it to predict. A prediction is the model’s estimate for a particular example.

A home-rent example

Suppose you want to estimate a home's monthly rent. You have details about its floor area, location, and number of bedrooms. Those details are inputs, often called features. The monthly rent is the target: the value you want to estimate.

What happens during training?

For supervised learning, past examples pair inputs with known outcomes. A model might see a home's size and location alongside its actual rent. Training adjusts the model to reduce errors between its predictions and those known rents.

For a new home, you provide the available inputs and ask the trained model for an estimate. You do not provide the unknown rent as an input—that is the value you are trying to predict.

Targets are not always numbers

Predicting rent is a regression task because the output is a numeric value. Predicting whether a photo contains a cat or a dog is a classification task because the output is a category. In both cases, the target gives the learning task a specific purpose.

Not every machine learning method uses labeled targets. Unsupervised learning, for example, can explore structure in data without a labeled outcome for every example. This guide focuses on supervised prediction.

How do you know whether the estimate is useful?

Evaluate the model on suitable examples it did not train on. Compare predictions with actual outcomes, and choose a measure that fits the decision. For rent, you might examine average absolute error in rupees and whether errors differ across locations or types of home.

Compare against a simple baseline, such as the average rent for similar homes. A model that is more complicated should earn that complexity by being more useful. A small average error can also hide large errors on individual homes, so inspect the distribution of mistakes.

Watch for clues you would not really have

A model can appear excellent if training inputs accidentally reveal the answer. Using a document created only after rent is agreed would be misleading if that document is unavailable when a real estimate is needed. This is one form of data leakage.

Ask whether each input will be available at the actual prediction time. Also check whether your evaluation examples represent future cases: a rent model trained in one city or year may perform poorly in another.

A prediction needs context

The model estimates a value; it does not establish a home's correct price or make a guarantee. Current market conditions, missing details, and unusual cases can change the result. Use predictions as evidence within a decision, with an appropriate check before acting.