Classification vs regression: losses and metrics

Look at what kind of value the model outputs.

Model answer. Both are supervised learning — the model learns from labelled examples — but they differ in the type of target.

Regression: predict a continuous number

• Examples: house price, tomorrow's temperature, a student's expected exam score.
• Training losses:
• Mean squared error (MSE): average of squared errors; penalises large errors heavily.
• Mean absolute error (MAE): less sensitive to outliers.
• Huber loss: behaves like MSE for small errors and MAE for large ones.
• Evaluation metrics: RMSE (same units as the target), MAE, and R² (fraction of variance explained relative to predicting the mean).

Classification: predict a category

• Examples: spam vs not spam (binary), digit 0–9 (multi-class), tags on a photo (multi-label).
• Training loss: cross-entropy (log loss), which compares predicted probabilities with the true label and heavily penalises confident wrong predictions. Models usually output probabilities via a sigmoid (binary) or softmax (multi-class).
• Evaluation metrics: accuracy, precision, recall, F1 score, ROC-AUC, and the confusion matrix. With imbalanced classes, accuracy alone is misleading.

Is logistic regression regression?
No — despite the name, it is a classification algorithm. It computes a linear combination of features and passes it through the sigmoid to output a probability between 0 and 1; a threshold (often 0.5) turns that into a class. The "regression" in the name refers to modelling the log-odds as a linear function of the inputs.

Some algorithms do both: decision trees, random forests, gradient boosting, k-nearest neighbours and neural networks have classification and regression versions; only the output layer, loss and leaf/aggregation rule change.

Common follow-up: "Predicting a movie rating from 1 to 5 stars — classification or regression?" It is ordinal: you could treat it as regression (and round), as classification (losing the order information), or use ordinal regression methods. A good answer explains the trade-off rather than picking one blindly.