Do preferences need numerical ratings?

No. Pairwise comparisons and rankings can be used directly by suitable methods.

Is preference learning the same as recommendation?

Recommendation is an application area. Preference learning is one family of techniques that can support it.

Can two people disagree consistently?

Yes. They may value different attributes. A useful model may need to represent that heterogeneity instead of treating all disagreement as random noise.

AI Glossary / AI and machine learning / Advanced

Preference learning

Definition

Preference learning uses observed or stated preferences to learn how alternatives should be compared, ranked or selected. Training information can include pairwise choices, ordered lists, ratings or behavior interpreted as evidence about what a person or group prefers.

What it means

It is often easier to say “I prefer A to B” than to assign each option an absolute score. Preference learning turns such relative judgments into a model that can help compare new alternatives.

The target may be one person's preferences, a group's aggregate ranking or a context-dependent choice. These are different goals. A model trained on a group average does not necessarily represent any particular individual, and disagreement can reflect real differences rather than incorrect annotation.

How it works

The team defines the alternatives, the context and the meaning of a preference label. A model then learns a scoring function or comparison rule consistent with the observed information, allowing for noise or uncertainty where appropriate.

Evaluation uses held-out comparisons or rankings relevant to the intended task. If preferences are inferred from clicks or purchases, exposure and presentation effects need attention: an item cannot be chosen if it was never shown. Data collection, model assumptions and the decision being supported all influence how the learned preferences should be interpreted.

Three practical examples

1. A search team can learn which results users find more useful, while accounting for the fact that highly placed results receive more attention.

2. A design group can collect comparisons between prototype layouts and investigate whether different user groups favor different features.

3. An AI-development team can use human comparisons of responses as one source of training feedback. Preference alone should not be assumed to verify factual correctness or suitability for every use.

Think of it this way

Imagine a librarian learning your taste by asking which of two books you would rather read. After several comparisons, the librarian can make more informed suggestions without requiring you to score every book in the catalog.

A computational model has a narrower basis for inference. It sees recorded comparisons and features, and may miss why your preference changes with mood, purpose or information unavailable in the data.

Worked example

Suppose a toy scoring model assigns A a score of 2, B a score of 1 and C a score of 0. It ranks A above B above C. Four held-out comparisons report A over B, A over C, B over C and B over A.

The ranking agrees with three of the four observations, giving pairwise agreement of 3/4 = 75%. The conflicting A-versus-B observations might come from different people or contexts; they do not automatically mean one label is wrong.

This simple metric does not reveal whether the most important comparisons were correct or whether a minority group was poorly served. Evaluation should reflect the actual decision and data-collection design.

Illustrative case study

Fictional learning platform SkillPath asks users to compare pairs of course descriptions. A model trained on all responses favors short introductory courses. The product team initially considers using that ranking for everyone.

Further analysis shows that experienced users prefer deeper material, while new users favor introductions. The team evaluates context-aware recommendations and keeps clear choices available rather than treating the aggregate as a universal preference. The case illustrates the difference between learning a population pattern and understanding each individual's needs; the numbers and outcome are not reported company results.

Common misconceptions and mistakes

A learned preference is not objective truth. A response can be attractive yet inaccurate, and a popular option can be unsuitable for a particular user. Distinguish preference prediction from factual verification and decision quality.

Another mistake is assuming clicks directly reveal liking without considering what was displayed and where. Preferences can also change over time. Avoid combining incomparable contexts, discarding disagreement automatically or reporting only an overall agreement score. A total ranking may be a useful simplification, but real preferences can involve ties, uncertainty and context-dependent reversals.

Business use

Preference learning can support recommendations, ranking and product research when comparison data is meaningful and representative. Start by specifying whose preferences matter and what the resulting ranking will influence. That helps determine the data, model and evaluation.

Monitor performance across relevant user groups and contexts. Give users ways to correct or refine personalization where appropriate. For consequential choices, keep other requirements separate from preference scores: affordability, factual accuracy or mandatory eligibility may need their own checks. A preference model supplies one input to a decision rather than defining the entire decision policy.

Check your understanding

An item receives few clicks after appearing near the bottom of a page. Can you conclude that users dislike it?

Answer: No. They may not have seen it. Exposure and position can affect observed behavior, so the data does not isolate preference without further assumptions or a suitable study design.

Related terms

Learning to rank
Utility function
Recommender System
Preference elicitation
pairwise comparison
reward model
personalization
position bias

Learn next

Collect a small set of comparisons and examine where they disagree. Separate differences between people from differences in context. Then compare pairwise prediction accuracy with the usefulness of the final ranked list for a specific task.

Sources

Thorsten Joachims, Optimizing Search Engines Using Clickthrough Data: https://www.cs.cornell.edu/people/tj/publications/joachims_02c.pdf. The Preference Learning Toolbox: https://arxiv.org/abs/1506.01709. Learning Linear Utility Functions From Pairwise Comparison Queries: https://arxiv.org/abs/2405.02612. Joachims's paper was consulted on relative preference feedback and the influence of which ranked links users see; the additional papers are further reading. The four-comparison calculation and SkillPath case are original illustrations.

Previous: Preference independence Next: Preference regression