Linear regression is a supervised learning algorithm. It requires labeled data to train, meaning every example in the training set must include both the input features and the correct output value. The model learns by comparing its predictions against these known answers and adjusting until the gap between them is as small as possible.
What Makes It Supervised
Supervised learning gets its name from the idea that the training data acts like a teacher. Each data point comes with a “ground truth” label that tells the model what the correct answer should be. Linear regression fits this definition exactly: you feed it a dataset where both the input variables (called predictors or independent variables) and the output variable (called the dependent variable) are known, and the algorithm finds the line or equation that best maps one to the other.
The output in linear regression is always a continuous number. You might use it to predict a salesperson’s yearly revenue based on their age and experience, estimate insurance claim costs from property data, or forecast how many games a basketball team will win based on points scored per game. In every case, you already have historical examples with known outcomes, and the model learns the pattern connecting inputs to those outcomes.
How the Model Learns From Labels
During training, linear regression makes a prediction for each example and then measures how far off it was from the actual value. This gap is called the loss. The most common way to measure it is mean squared error (MSE), which takes each prediction error, squares it, and averages them all together. Squaring does two useful things: it removes negative signs so errors don’t cancel each other out, and it penalizes large mistakes more heavily than small ones.
The model then adjusts its internal parameters to shrink that loss. It keeps iterating until the predictions are as close to the real values as possible. This entire process depends on having labeled data. Without the actual values to compare against, the model would have no way to know whether its predictions were good or bad, and no direction to improve.
How Unsupervised Learning Differs
Unsupervised learning works with unlabeled data. There are no correct answers provided. Instead, the algorithm looks for hidden structure in the data on its own. A common example is k-means clustering, which groups data points into clusters based on similarity. The algorithm finds natural groupings by minimizing the distance between points within each cluster, but nobody tells it what those clusters should be or what they mean.
The core difference comes down to the goal. Linear regression tries to predict a specific known outcome. Clustering tries to discover patterns nobody has defined yet. One needs a teacher; the other doesn’t.
Where Linear Regression Fits Among Other Algorithms
Supervised learning splits into two main categories: regression and classification. Linear regression falls on the regression side because it predicts continuous numerical values. Its close relative, logistic regression, falls on the classification side because it predicts categories (like yes/no or spam/not spam), even though the name sounds similar.
Other supervised regression algorithms include ridge regression and lasso regression, which are variations of linear regression that add penalties to prevent the model from fitting too closely to the training data. Polynomial regression extends linear regression to capture curved relationships. All of them share the same fundamental requirement: labeled data with known output values.
How You Know If It’s Working
Because linear regression is supervised, you can measure exactly how well it performs by comparing predictions to actual values. The most commonly used metrics include:
- R-squared: Tells you what proportion of the variation in your outcome the model explains. A perfect score is 1.0, meaning the model accounts for all the variation. A score near 0 means it explains almost nothing.
- Mean squared error (MSE): The average of all squared prediction errors. Lower is better, with 0 being perfect.
- Root mean squared error (RMSE): The square root of MSE, which puts the error back into the same units as your original data, making it easier to interpret.
- Mean absolute error (MAE): The average of the absolute differences between predictions and actual values. Less sensitive to outliers than MSE because it doesn’t square the errors.
None of these metrics would be possible without labeled data. Each one requires you to compare a predicted value against a known correct answer. This is the signature of supervised learning, and it’s baked into every step of how linear regression is built, trained, and evaluated.