Diskriminativ

Diskriminativ is a term used in statistics and machine learning to describe models and learning methods that focus on predicting the label Y from input X, rather than modeling the full data distribution. In this sense, discriminative approaches aim to model the conditional distribution P(Y|X) directly or to learn a decision boundary between classes, without necessarily specifying how the data X were generated.

Compared with generative models, which model P(X|Y) and P(Y) and then apply Bayes' rule to obtain P(Y|X), discriminative models typically require fewer assumptions about the joint distribution. They often yield higher predictive accuracy when the goal is classification, especially with limited data, because they concentrate learning capacity on the discrimination task rather than on data generation.

Common discriminative methods include logistic regression, linear and nonlinear support-vector machines, and many neural networks trained for classification. Conditional random fields are another class of discriminative models used for structured prediction. The term can also describe training objectives, such as cross-entropy or hinge loss, that optimize P(Y|X) or a margin between classes.

Advantages of discriminative models include strong predictive performance and flexible modeling of complex decision boundaries. Limitations include weaker ability to generate new samples or to model the underlying data distribution, which can be relevant for tasks such as anomaly detection or unsupervised learning. In modern practice, discriminative models are widely used for supervised classification and structured prediction, often in combination with generative approaches in hybrid methods.