Context Is Part of the Preference

A person may be strongly attracted to an object and still decide against it. The object may be inappropriate for the occasion, outside a temporary budget, intended for someone else, or simply wrong for the task at hand. Conversely, an object that would normally attract little interest may become highly relevant under a particular set of circumstances.

Neither case necessarily requires an underlying change in taste. What changes is the relationship between a relatively persistent set of tendencies and the situation in which a decision has to be made.

This distinction matters for AI systems because preference is rarely available as a directly observable variable. Systems encounter evidence: searches, choices, ratings, purchases, rejections, descriptions and other interactions. To infer a person’s taste from those observations, a model has to decide which aspects are likely to generalize beyond the situation that produced them.

Removing context is therefore not simply a way of removing noise. It can alter what the observation appears to say about the person.

Established research

Preference can depend on how it is elicited

A substantial literature in behavioral decision research has challenged the assumption that people always approach a decision with a complete and invariant ordering of the available alternatives.

Slovic’s work on the construction of preference reviewed experiments in which normatively equivalent elicitation procedures produced systematically different responses. In classic preference-reversal studies, for example, choices between risky prospects did not always imply the same ordering as monetary valuations of those prospects. These findings motivated an account in which at least some preferences are constructed during judgment rather than simply retrieved in finished form (Slovic, 1995).

Bettman, Luce and Payne developed this idea in consumer decision making. Their account of constructive choice emphasizes that people have limited processing capacity and often adapt their decision strategies to the demands of the task, the information available and the goals they are pursuing (Bettman, Luce & Payne, 1998).

It would be easy to overread this literature as showing that preferences are fundamentally unstable. That is not the conclusion the same researchers draw. Bettman, Luce and Payne later argued explicitly that preference construction and preference stability are compatible: some preferences may remain relatively persistent even when their expression or the processes used to arrive at a choice vary with context (Bettman, Luce & Payne, 2008).

For the problem of representing taste, this coexistence is more useful than either extreme. A person need not be represented as possessing one invariant ranking over everything they might encounter, but neither must every choice be treated as a new and unrelated preference.

The available alternatives can affect evaluation

Another line of research concerns the composition of the choice set itself. Tversky and Simonson showed that the relative attractiveness of alternatives can depend on the other options with which they are compared, challenging models in which every option carries a value that is independent of its local choice context (Tversky & Simonson, 1993).

Related work on attraction, compromise and similarity effects produced numerous demonstrations of such context dependence. These effects should not, however, be treated as universal laws of choice. Later research has documented substantial boundary conditions and cases in which small changes to an experimental setting weaken, eliminate or reverse them (Spektor, Bhatia & Gluth, 2021).

The conservative conclusion is enough for our purposes. Human evaluation is not guaranteed to remain invariant when the surrounding decision problem changes. An interaction with an object can therefore contain information not only about the object and the person, but also about the comparison in which the interaction occurred.

Goals change the decision space

Context is broader than the alternatives visible at one moment. What a person is trying to achieve can change which objects are considered relevant in the first place.

Research on goal-derived categories has shown that consideration sets can cross conventional product categories when people organize alternatives around a salient goal rather than around the taxonomy of the products themselves (Ratneshwar, Pechmann & Shocker, 1996). In practical terms, “things I might buy” and “things that could solve this particular problem” are not necessarily the same set.

The identity of the beneficiary also changes the decision. A review of consumer choices made for others distinguishes situations such as gift-giving, joint consumption, favors and caregiving, each of which can involve a different balance between the chooser’s preferences, the recipient’s preferences and the relationship between them (Liu, Dallas & Fitzsimons, 2019).

This matters because an observed action does not identify its own objective. The purchase of an object may express the purchaser’s taste, an attempt to predict someone else’s taste, a compromise between several people, or a practical response to a particular requirement. Treating all four as equivalent preference observations introduces assumptions that are not present in the behavior itself.

Recommender systems have long treated context as a distinct dimension

The computational literature reached a related problem through recommendation. Traditional formulations often centered on users, items and interactions between them. Context-aware recommender systems extended this formulation by incorporating additional information such as time, location, activity or social setting (Adomavicius et al., 2005; Adomavicius et al., 2011).

This field has produced several ways of incorporating context: selecting contextually relevant observations before recommendation, adjusting recommendations after an initial prediction, or incorporating contextual information directly into the predictive model. More recent work has extended these approaches through factorization methods, neural architectures and other representation-learning techniques.

Yet the definition of context remains less straightforward than the terminology suggests. A recent systematic review of context-aware recommender systems distinguishes between a representational view, in which context consists of predefined and potentially observable variables, and an interactional view, in which context can emerge through relationships between activities. It also emphasizes that psychologically important factors such as mood or intent may be only partially observable or entirely latent (Mateos & Bellogín, 2025).

This distinction is consequential. Time, location and device can often be recorded directly. “Preparing for an important meeting,” “looking for something unfamiliar,” or “choosing for a close friend” are higher-level situations whose relevance depends on an interpretation of what the person is doing. Recent work on situation-aware recommender systems makes a similar distinction between raw contextual conditions and higher-level representations intended to capture the state, goals and likely evolution of a situation (Aliberti, D’Aniello & Gaeta, 2026).

The computational problem is therefore not solved simply by attaching a longer list of contextual features to every interaction.

Interpretation

An observation has a scope

Suppose a system observes that someone spent several minutes examining a particular watch. The observation is real, but several interpretations remain possible. The person may be considering buying it, researching a gift, comparing one technical characteristic with another model, studying the design without wanting to own it, or evaluating it for some entirely different purpose.

The ambiguity cannot be resolved from the object alone. Nor can it necessarily be resolved by accumulating more observations of the same type. What matters is the relationship between the behavior and the situation in which it occurred.

This suggests a useful distinction between the strength of preference evidence and its scope. An observation may be strong evidence within one class of situations while providing weak evidence outside it. Choosing a jacket for oneself may tell us considerably more about one’s own clothing preferences than choosing a jacket for a sibling. Rejecting an expensive chair during a temporary budget constraint may provide good evidence about the current decision while providing very little evidence that the person dislikes the chair.

In each case, the visible behavior is clear. The uncertainty concerns what should be generalized from it.

Context therefore affects the interpretation of evidence before it affects the quality of a future recommendation. Once an observation has been reduced to an unconditional proposition such as “likes formal clothing” or “dislikes this designer,” information about the conditions under which that inference was justified may already have been lost.

Context should not become a synonym for everything

There is a risk in making the concept too broad. If context includes every fact that can possibly influence behavior, it stops being analytically useful.

It is helpful to separate several kinds of information. Environmental conditions include variables such as location, time and physical setting. Decision context concerns the alternatives available and the way the decision is posed. Social context concerns the people involved and their relationships. Goal context concerns what the person is trying to accomplish. Constraints describe conditions that restrict the available action without necessarily changing what the person prefers.

These categories can interact, and the boundaries between them are imperfect. The distinction is nevertheless important because they have different implications for what can be inferred about a person.

A budget constraint, for example, is not itself an aesthetic preference. A gift recipient’s taste does not become the giver’s taste. A dress code does not determine what the wearer finds beautiful. If an AI system collapses all of these influences into a single notion of preference, it risks attributing to the person what actually belonged to the circumstances.

Paradoxically, context is useful for modelling persistent taste partly because it helps identify what should not be treated as persistent taste.

TasteGraph perspective

We think the usual distinction between a stable user model and a contextual recommendation problem is useful but incomplete.

Context is often introduced at the moment a system has to decide what is relevant now. A persistent representation describes the person; current contextual information then modifies what the system retrieves, ranks or recommends.

There is another question earlier in the process: under what conditions was the evidence used to construct that persistent representation generated?

The issue is not primarily historical. It is about conditional meaning. If an interaction occurred while choosing for someone else, responding to a constraint or pursuing an unusual goal, that condition affects how strongly the observation should generalize to other situations. A model can know the context of today’s decision and still misinterpret the person if the evidence from which its representation was learned has already been stripped of those distinctions.

This is why we do not see context as the opposite of persistent structure.

People appear to have recurring tendencies. They return to particular forms, materials, genres, proportions, flavors, makers and ideas. Without some persistence of this kind, the concept of taste would have little explanatory value. But persistent tendencies need not be expressed uniformly across every situation.

Someone may consistently prefer restrained interiors and still choose an exuberant restaurant. They may generally avoid conspicuous branding and nevertheless value one object whose history makes the branding meaningful. They may like experimental music and prefer silence while working. None of these observations requires taste to become arbitrary; each suggests that different parts of a person’s preference structure become relevant under different conditions.

This also helps clarify what we mean when we say that context is part of the preference.

We do not mean that a constraint is itself a taste, or that every situational influence should become part of a person’s identity. We mean that, when an AI system uses behavior as evidence of taste, the context in which that behavior occurs can be necessary to interpret what the evidence says and where it should generalize.

The object alone is insufficient. The action alone may also be insufficient.

A useful representation has to preserve enough of the relationship between person, object and situation to distinguish persistent preference from conditional relevance.

Otherwise, simplifying away context can simplify away part of the information the representation was intended to capture.