What are features?
Features are the input variables a machine learning model uses to make a prediction. In a table, each feature is usually a column and each row is one example, such as one customer, one invoice, or one sensor reading. The value the model predicts is the target, or label, and it is kept apart from the features. A model can only find patterns in the inputs it receives, so a fact left out of the features cannot inform the prediction.
Types of features
- Numeric features hold quantities such as a balance, a weight, or a temperature.
- Categorical features hold one value from a fixed set, such as a state or a product line, and are usually converted to numbers before training.
- Derived features are calculated from raw data, such as the days since a customer’s last order, and building them is the work of feature engineering.
- Extracted features turn text, images, or audio into numbers, often as embeddings, through feature extraction.
Features in practice
A collections agency ranks accounts by likelihood of payment so collectors call the most promising accounts first. Candidate features include the balance, the age of the debt, the number of past payments, whether a working phone number is on file, and the outcome of the last contact. Feature importance measures then show which of these carry the most signal.
Review also catches problem features. A field filled in only after an account pays would make the model look accurate in testing and fail in production, a mistake called information leakage. A field such as ZIP code can stand in for protected characteristics such as race, so it needs a fairness review before use.
What to watch for
- Every feature must be available at prediction time, with the value it had at that moment.
- Define each feature once and reuse the definition in training and production, for example through a feature store.
- Irrelevant features add noise and raise the risk of overfitting, so more features do not guarantee a better model.
- Document the source and meaning of each feature so auditors and new analysts can trace what the model saw.