Discriminative Model

A discriminative model is a type of machine learning model that learns to distinguish between categories or predict a target variable directly from input features, without modeling the underlying data distribution that generates those inputs. Rather than asking "how likely is this kind of data to exist?" (as generative models do), a discriminative model asks "given this data, what is the most likely label or value?" Common examples include logistic regression, support vector machines, and classification-focused neural networks. These models learn a decision boundary or mapping function optimized for the specific prediction task.

In commerce and enterprise AI, discriminative models are the workhorses of classification and prediction tasks: determining whether a transaction is fraudulent or legitimate, whether a customer is likely to churn, whether a product review is positive or negative, or whether a search query should surface a particular product. They are typically more sample-efficient and computationally economical than generative alternatives for these tasks because they focus all their modeling capacity on the prediction boundary rather than reconstructing the full data distribution. Understanding the discriminative versus generative distinction helps AI practitioners select the right model architecture for a given problem—classification and ranking tasks favor discriminative approaches, while content creation, data augmentation, and simulation tasks favor generative ones.