MMOOZZGG.com

Glossary

What is a discriminative model?

In short

A discriminative model is a machine learning model that predicts an answer, such as a class, directly from the features of an object and does not try to describe how the data itself is built.

AI basics · discriminative model

MMOOZZGG.com editorsChecked: 1 min read

For a child

A teacher shows you two piles of pebbles: smooth ones and rough ones. You do not need to know how the pebbles got into the river or what they are made of. You just run a finger over a new one and say which pile it goes into. Some programs learn in a similar way: they do not draw or write stories, they only decide fast where a thing belongs.

In plain words

Machine learning models fall into two camps. Generative models learn what the data looks like, so they can create new data: texts, pictures, voice. Discriminative models only learn to draw a border between possible answers. A mail filter looks at an email and decides whether it is spam; it does not need to write emails. A bank system looks at an application and estimates the risk of default. Such models are simpler, faster and often more accurate at their narrow task, but they do not create new data.

For business

Most of the AI that quietly makes money works this way: credit scoring, fraud detection, request classification, defect detection on a production line, churn forecasts. In operation such models cost less than generative ones, answer in fractions of a second and give a clear quality metric: precision, recall, error rate. They need labeled examples, and labeling often sets the cost of the project. The risk is in the data: if the training set is skewed, the model repeats the skew, for example by refusing a certain group of clients more often. Before launch, quality is checked by segment, not only on average. When incoming requests are processed, such a classifier is often the first link, before the reply and the hand-off to the CRM.

For a specialist

A discriminative model estimates the conditional distribution P(y|x) directly or builds a decision function f(x) without modeling the distribution of features. A generative model estimates the joint P(x, y) or P(x) and gets P(y|x) by Bayes’ rule. Typical discriminative methods are logistic regression, support vector machines (SVM), conditional random fields (CRF), gradient boosting and most neural network classifiers. A. Ng and M. Jordan (NIPS 2001) compared logistic regression with the naive Bayes classifier and showed two regimes: the asymptotic error of the discriminative method is lower, but the generative one approaches its own, higher asymptote faster, so it can win on small samples. In generative adversarial networks (GANs), the discriminator is the part that tells real examples from generated ones: a special case of the same approach.

Do not confuse. Do not confuse it with discrimination against people: here the word means “to tell apart”, that is, to separate one class of objects from another.

Example

An insurance company trains a model on past claims, and the model estimates the probability of fraud from twenty features. Suspicious cases go to a specialist, and the rest are paid automatically.

Related terms

Questions and answers

How does a discriminative model differ from a generative one? A discriminative model learns to tell classes apart and answers the question “what is this”. A generative model learns how the data is built and can create new examples.

For classification the first is usually enough; for texts and images you need the second.

Which models are discriminative? Logistic regression, support vector machines, gradient boosting and most neural network classifiers. They are used for credit scoring, spam filtering and object recognition.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us