Skip to content

Common Loss Functions

A loss function (or cost function) is the mathematical equation an algorithm uses to calculate how "wrong" its predictions are. The algorithm then adjusts its parameters to minimise this value.

Regression Loss Functions

Used by algorithms predicting continuous values (e.g., LinearRegression, Ridge).

Mean Squared Error (MSE)

\[\text{MSE} = \frac{1}{n} \sum_{i=1}^{n} (y_i - \hat{y}_i)^2\]
  • Heavily punishes large errors because the difference is squared.
  • Sensitive to outliers — a single extreme prediction dominates the total.
  • Use when: Large errors are disproportionately costly (e.g., financial forecasting).

Mean Absolute Error (MAE)

\[\text{MAE} = \frac{1}{n} \sum_{i=1}^{n} |y_i - \hat{y}_i|\]
  • Treats all errors equally regardless of magnitude.
  • More robust to outliers than MSE.
  • Use when: You want a metric that reflects typical error magnitude without outlier distortion.

Root Mean Squared Error (RMSE)

\[\text{RMSE} = \sqrt{\text{MSE}}\]
  • Same units as the target variable, making it easier to interpret than MSE.
  • Still penalises large errors heavily due to the underlying squaring.

Classification Loss Functions

Used by algorithms predicting discrete categories (e.g., LogisticRegression, RandomForestClassifier).

Log Loss (Binary Cross-Entropy)

\[\text{Log Loss} = -\frac{1}{n} \sum [y \log(\hat{p}) + (1 - y) \log(1 - \hat{p})]\]
  • Punishes confident but incorrect predictions severely (e.g., predicting 99% probability of Class 1 when the true label is Class 0).
  • Use when: You need well-calibrated probability outputs, not just hard class labels.

Hinge Loss

\[\text{Hinge} = \max(0, 1 - y \cdot \hat{y})\]
  • Used by Support Vector Machines (SVM).
  • Penalises misclassifications and observations that fall within the margin, but incurs zero loss for correctly classified points far from the boundary.

Quick Reference Table

Loss Function Problem Type Outlier Sensitivity scikit-learn Parameter
MSE Regression High scoring="neg_mean_squared_error"
MAE Regression Low scoring="neg_mean_absolute_error"
Log Loss Classification Medium scoring="neg_log_loss"
Hinge Classification Low scoring="hinge" (SVM)

KSB Mapping

KSB Description How This Addresses It
K4.1 Statistical models and methods Understanding the statistical basis of regression and classification
K4.2 ML and AI techniques Implementing and comparing supervised learning algorithms
K4.4 Resource constraints and trade-offs Model complexity vs interpretability; computational cost
S1 Scientific methods and hypothesis testing Formulating hypotheses and testing with rigorous validation
S4 Building models and validating Cross-validation, train/test evaluation, performance metrics
B5 Impartial, hypothesis-driven approach Honest evaluation of model performance and limitations