How it works
A feature, or attribute, is a characteristic of a sample that the machine learning model uses as input. An image, a message, or a customer can be represented by a set of features that describe what is being analyzed. Each sample is transformed into a vector of numbers that captures relevant aspects, and the model learns to combine these values to produce a prediction. If the features are weak, the model cannot generalize, even with large amounts of data. Feature quality determines the upper performance limit of any model, regardless of its architecture. Values are typically standardized or normalized before entering the network, so that a feature with a huge range does not dominate the others. Selecting good features requires cross-validation and continuous monitoring, because real-world behavior changes and the model may need new inputs.
Practical example
To predict whether a customer will buy, features can include age, purchase history, time since last visit, and product type. The model combines these numbers and outputs a conversion probability. If the features are noisy, the model may look good on historical data and break in production. Data engineers validate features with statistics and track changes in customer behavior before investing in a model. Even a highly complex model produces weak results when the input representation is poorly designed. If a customer does not buy, the features may indicate that the ad did not reach the right audience or that the price was outside the buyer's expectations.
What it's for
Feature engineering determines whether the model can learn useful patterns. Good features reduce the amount of data needed and improve accuracy. Bad features produce models that memorize noise and break in production. The feature system also makes it easier for business and engineering teams to use the same data definition, reducing discrepancies between what the model learns and what the dashboard shows. In real projects, feature selection starts with a list of business factors, and each factor is checked to see if it represents something the model can learn with confidence.
Difference between X and Y
A feature is the input representation; a label is the output the model tries to predict. A feature store is the system that stores and shares these representations across engineering and data science teams.