Must the system learn a complete utility function?

No. Partial information can be enough to support a particular choice.

Are direct numerical ratings required?

No. Comparisons and rankings can also provide useful information.

Can observed behavior reveal preferences?

It can provide evidence, but availability, context and other constraints may affect what a person does.

AI Glossary / Graphs and symbolic AI / Advanced

Preference elicitation

Definition

Preference elicitation is the process of obtaining information about what a person or group prefers so a system can support decisions on their behalf. It can involve comparisons, rankings, trade-off questions or other responses used to construct or refine a preference model. The aim may be to identify a useful choice without determining every detail of a complete utility function.

What it means

Asking whether an earlier delivery is worth a higher price supplies information about that user's priorities. One answer usually rules out some preference models rather than uniquely identifying the correct model. A system should distinguish what the user actually stated from conclusions inferred under those assumptions.

How it works

Specify the decision alternatives and attributes that matter, choose a preference representation, and ask questions that help distinguish relevant alternatives. Translate each response into information about the model. For example, a pairwise choice can constrain the relative weights on two normalized attributes.
Select further questions according to what remains uncertain and what could change the recommendation. In regret-based approaches, the system can evaluate how much utility might be lost by a choice under remaining plausible models, then seek information that reduces that uncertainty. The process may stop when the decision is sufficiently supported, when additional questions offer little value or when the user does not wish to continue. Its conclusions depend on the preference model, the available alternatives and the reliability of the responses.

Three practical examples

A delivery assistant asks whether the customer prefers a cheaper arrival window or a more expensive earlier one. The answer helps clarify a trade-off instead of assuming that speed is always more important.
A product configurator asks users to compare complete configurations. This can reveal interactions that a checklist of independently preferred features would miss.
A meeting assistant asks which of two feasible schedules is less inconvenient. It can use the answer to refine an explicit model of scheduling preferences, while recognizing that one participant's answer does not automatically represent everyone's priorities.

Think of it this way

Imagine a helpful shop assistant who narrows the options by asking meaningful comparisons. The assistant does not need your exact numerical satisfaction with every item to help you choose. But asking only about color would be inadequate if size and durability determine your decision.

Worked example

Suppose three service options have two scores: affordability and speed, each measured on an agreed scale from 0 to 10, with higher scores better. Option A has scores (9, 3), B has (5, 8), and C has (7, 6).
For illustration, assume the user's utility follows U = w times affordability + (1 - w) times speed, where 0 is less than or equal to w and w is less than or equal to 1. This additive model and the score scales are assumptions, not facts established by asking a single question.
The utilities are U(A) = 3 + 6w, U(B) = 8 - 3w, and U(C) = 6 + w. Initially the system does not know w. If the user says C is strictly preferred to B, the model requires 6 + w > 8 - 3w. Hence 4w > 2, or w > 0.5.
A second answer says C is strictly preferred to A. That requires 6 + w > 3 + 6w, giving 3 > 5w, or w < 0.6. Together the answers restrict w to the open interval (0.5, 0.6).
The exact weight remains unknown, but C is the uniquely preferred option among these three for every weight in that interval. At w = 0.55, for example, the utilities are 6.30 for A, 6.35 for B and 6.55 for C. The system can recommend C among the stated alternatives without claiming to have recovered a precise weight.
Now add option D with scores (8, 5). Its utility is 5 + 3w. The difference U(D) - U(C) is 2w - 1, which is positive throughout the remaining interval. D therefore outranks C there. The earlier recommendation was supported for the original three options; it was not a guarantee against alternatives that had not yet been considered.

Illustrative case study

In a fictional service-selection pilot, a team asks users to rate 'importance of price' on an undefined scale and inserts the answer directly as a weight.
The team instead presents the concrete comparisons above, describes the attribute scales and records the constraints implied by each response. It reports a range of compatible weights rather than inventing a precise estimate. When a new service appears, it reevaluates the recommendation. It also asks users whether affordability and speed capture the decision adequately; a missing requirement, such as a needed feature, could make the two-attribute model inappropriate regardless of the arithmetic.

Common misconceptions and mistakes

Eliciting a preference is not the same as discovering an objective product quality. Different users can reasonably rank the same options differently. A small set of choices also need not identify a unique utility function.
Do not interpret indifference, uncertainty and refusal to answer as interchangeable. Inconsistent responses may indicate noise, changing context or a poor model, rather than justify silently discarding an inconvenient answer. Finally, learned weights depend on the attribute scales: changing a score's units without adjusting the model can change recommendations unintentionally.

Business use

Preference elicitation can make recommendations more relevant and explainable when important trade-offs are uncertain. Ask questions that affect the decision, show understandable alternatives and allow correction. Separate mandatory requirements from negotiable preferences. Record the assumptions behind inferred weights or rankings and revisit them when options or circumstances change. For group decisions, specify whose preferences are being gathered and how they are combined. Evaluate the interaction's burden alongside recommendation quality; a long questionnaire can be costly even when its mathematical model is sophisticated.

Check your understanding

What weight range follows from the two strict comparisons?

Answer: Greater than 0.5 and less than 0.6.

Does that determine an exact weight?

Answer: No.

Why must the recommendation be reconsidered when D appears?

Answer: The original conclusion compared only A, B and C.

Related terms

Utility function
Preference constraints
Partial information
preference model
pairwise comparison
multiattribute utility
minimax regret
recommendation
trade-off

Learn next

Replace the first strict preference with indifference between C and B. Solve for w and compare all four options, explicitly allowing tied rankings.

Sources

Boutilier and colleagues, Constraint-based Optimization and Utility Elicitation using the Minimax Decision Criterion, author abstract: https://www.cs.toronto.edu/~cebly/Papers/BPPS-aij06-abs.html Poole and Mackworth, Preferences and Utility: https://artint.info/3e/html/ArtInt3e.Ch12.S1.html AIMA glossary, preference elicitation: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Preference constraints Next: Preference independence