Discriminant analysis
Classify observations into known groups
Discriminant analysis uses labelled observations to estimate rules for classification. Linear discriminant analysis assumes shared within-group covariance under its usual Gaussian formulation; quadratic discriminant analysis allows group-specific covariance. Mixture discriminant models permit more complex class distributions. This is supervised classification, distinct from clustering groups that have no prior labels.
Use when group labels are known and numeric predictors may distinguish them. Ensure representative labelled training data and evaluate classification on independent observations using the deployment class frequencies and costs.
Strengths
- Produces interpretable discriminant functions
- Can model classification probabilities under explicit distribution assumptions
Limitations
- Covariance estimation is fragile with many predictors
- Distribution and class-prior shifts can impair predictions
Know the boundary
High training accuracy does not demonstrate generalisation or explain why the groups differ.