Data, Features and Labels

Every machine learning project starts with data. A dataset is a collection of examples (also called samples or rows). Each example is described by:

• Features: the measurable inputs. For a house: size in square meters, number of bedrooms, postcode. For an image: its pixel values. For text: its tokens.
• A label (or target): the answer you want the model to predict, such as the sale price or “spam / not spam”.

Supervised learning learns the mapping from features to labels. Unsupervised learning only has features and looks for structure in them. Self-supervised learning invents labels from the data itself, for example “the next word”.

A few practical truths:

• Garbage in, garbage out. Mislabeled, duplicated or unrepresentative examples hurt more than a slightly worse model choice. Getting good labels is its own discipline: see data labeling.
• Features can be engineered. Turning a date into “day of week” or “is holiday” often helps a classical model a lot. See feature engineering.
• Deep learning learns its own features from raw inputs like pixels, which is a big reason it took over vision and language.
• Data must be split so the model is judged on examples it never saw. See train/test split.