← AI Terminology

Features

Features are the individual measurable properties or input variables fed into a machine learning model to make predictions.

They are the raw signals a model learns from — everything from pixel values and word counts to stock prices and sensor readings.
Why It Matters in AI
The quality and selection of features often determines model performance more than the choice of algorithm itself. A model can only learn patterns that exist in the features it receives; missing or irrelevant features create a ceiling on accuracy. Feature engineering — crafting and transforming raw data into informative inputs — remains one of the highest-leverage skills in applied ML.
Key Points
Aspect Description
Types Numerical (continuous/discrete), categorical (nominal/ordinal), text, image, time-series
Also called Predictors, input variables, covariates, attributes, columns
Common tools pandas for tabular prep, sklearn.preprocessing, sklearn.feature_selection
Feature selection Choosing the most informative subset (reduces noise and overfitting)
Feature engineering Manually creating or transforming features (e.g. log-scaling, one-hot encoding, lag variables)
Role in deep learning Raw features (pixels, tokens) are often used directly; the network learns its own representations
Simple Analogy
Think of features as the columns in a spreadsheet you hand to a hiring manager to evaluate job candidates. Height, GPA, years of experience — each column is one feature. The better the columns you choose (and the cleaner the data), the better the manager's judgment will be.
Common Usage Examples
  • X = df[['age', 'income', 'debt_ratio']] — selecting feature columns before model training
  • One-hot encoding a country column: pd.get_dummies(df['country'])
  • SelectKBest in scikit-learn to rank and filter features by statistical importance
  • In NLP: TF-IDF scores as features for a text classifier
  • In computer vision: raw pixel values or extracted embeddings as features for an image model
Summary
In short: Features are what a model sees — choose them well and half the modelling problem is already solved.