Machine Learning Terms: Features, Labels, Training, and Inference

Machine learning uses data to fit a model that can make a prediction or identify a pattern. The words feature, label, training, and inference describe different parts of that process. A house-price example makes them easier to distinguish.

Features and labels

In supervised learning, a feature is an input available to the model, such as a house's floor area or age. A label is the known outcome used during training, such as its recorded sale price. In the two comparisons below, algebra's x and y correspond to an input feature and an output label. The second column uses w for a learned weight and b for an intercept.

A simple linear model with one numeric feature can be written ŷ = b + wx, where ŷ means a predicted value and the observed label is y. With multiple features, the model can use a sum such as ŷ = b + w₁x₁ + w₂x₂. Many models are not linear, so these equations are illustrations rather than a universal definition.

Model, training, evaluation, and inference

A model is the fitted mapping from features to an output. Training adjusts its parameters using examples. Evaluation tests it on separate examples and compares predictions with actual labels. Inference applies the trained model to new inputs, such as estimating the price of a house that has not sold yet. An estimate can still be wrong, especially when new data differs from the training set.

Google's supervised learning introduction explains features, labels, and inference in more depth. TipsMake also explains how machine learning fits within AI.

Different learning setups

• Supervised learning: Examples include known labels, such as email marked spam or not spam.
• Unsupervised learning: The model finds structure without a target label, for example grouping similar items. A group is not automatically meaningful; someone must interpret it.
• Reinforcement learning: An agent interacts with an environment and receives a reward signal. This is a distinct setup rather than unsupervised learning with simple feedback.
• Self-supervised learning: Training targets come from the data itself, such as predicting a masked part of a sequence. The term does not mean the model independently monitors itself in production.

For related architectures, see the deep learning introduction. Google's machine learning glossary is a useful reference when terminology differs across courses.