What is a discriminative model?
A discriminative model is a machine learning model that learns to tell classes apart, or to map inputs directly to an output, without learning how the inputs themselves are distributed. In statistical terms it estimates the probability of a label given the input. Logistic regression, support vector machines, random forests, gradient-boosted trees, and most neural network classifiers are discriminative. They answer a narrow question well: given this claim, this transaction, or this image, which category does it belong to, and with what confidence.
Most scoring and classification in a business is discriminative. A churn score, a fraud flag, a defect detector on a production line, and a document router that sends invoices to accounts payable all learn a boundary between outcomes from labeled examples. They need supervised learning with labeled training data, and they cannot produce new examples of the data they classify.
Discriminative models compared with generative models
| Aspect | Discriminative model | Generative model |
|---|---|---|
| What it learns | The boundary between classes | The distribution of the data itself |
| Typical output | A label or a score | A new sample, such as text or an image |
| Data needed | Labeled examples | Large volumes of examples, labeled or not |
| Common examples | Logistic regression, gradient-boosted trees | Large language models, diffusion models |
A healthcare billing team that wants to predict which claims will be denied needs a discriminative model, and a small one trained on its own claims history usually beats a general-purpose language model on that task at far lower cost.