Taste Is Not a Vector

Machine learning has become exceptionally good at turning complex things into points in latent spaces.

Words, images, products, songs and users can all be mapped into numerical representations in which useful relationships become accessible to computation. In recommender systems, this idea has been especially productive. Matrix-factorization methods showed how observed interactions could be explained through latent user and item factors, while later neural systems learned increasingly expressive representations from much larger and more heterogeneous behavioural histories. [1][2]

There is no argument here against that approach. Embeddings are powerful precisely because they compress. They preserve information useful for a task while making comparison, retrieval and prediction manageable at scales where reasoning directly over every observation would often be impractical.

“Taste is not a vector” is therefore not intended as a mathematical claim about what a vector can or cannot encode. A sufficiently rich numerical representation can, in principle, contain a great deal of structure. Nor is it a claim that one representation is always inferior to several.

The question is about representational adequacy.

A user representation is normally learned or constructed for some purpose: retrieving likely candidates, estimating a click probability, ranking items, predicting a rating, or optimizing another downstream objective. Its usefulness is measured against that objective.

Success at one of those tasks establishes something important about the representation. It does not necessarily establish that the same representation preserves every distinction that would matter when reasoning more broadly about the person.

That difference is the subject of this note.

Established research

Recommender systems provide an unusually useful empirical setting for thinking about preference representation. This is not because recommendation defines the problem of taste, but because recommendation has forced computational models of preference to operate at enormous scale, against measurable outcomes, for several decades.

Early latent-factor approaches offered an elegant abstraction: users and items could occupy a shared latent space, and their compatibility could be estimated from their positions within it. The success of these methods demonstrated how much predictive structure could be recovered without explicitly specifying every dimension of preference. Deep recommendation systems retained the general advantages of learned representations while increasing their capacity and incorporating richer behavioural signals. [1][2]

At the same time, several influential architectures have found that a single, fixed user representation is not always the most useful abstraction even within relatively narrow recommendation tasks.

The Deep Interest Network developed at Alibaba, for example, was motivated by a limitation in architectures that compressed a user’s features into the same fixed-length representation regardless of the candidate advertisement. DIN instead computes a representation of the user’s relevant historical behaviour with respect to the candidate being evaluated. The user is still represented numerically, but the representation is conditional: different candidates can activate different parts of the same behavioural history. [3]

Other systems have made multiplicity explicit. MIND, developed for large-scale recommendation at Tmall, represents a user with multiple interest vectors rather than a single overall embedding. Its interest extractor groups patterns in historical behaviour into several representations that can participate separately in candidate matching. ComiRec follows a related premise in sequential recommendation, arguing that a unified representation of a behavioural sequence may obscure multiple interests active during the same period. [4][5]

Temporal models introduce another dimension. Koren’s work on temporal collaborative filtering showed that modelling changes in user preferences and item perception over time could materially improve recommendation. DIEN later modelled the evolution of latent interests through behavioural sequences and conditioned that evolution on the target item under consideration. [6][7]

These are not isolated observations. Multi-interest recommendation has developed into a substantial area of research; a 2026 survey in ACM Transactions on Information Systems characterizes the field around the difficulty of representing multifaceted preferences from diverse, volatile and sometimes ambiguous behavioural evidence. [8]

It would be easy to draw too much from this literature. DIN, MIND, DIEN and ComiRec are engineering systems built to improve particular recommendation or advertising objectives. Their empirical success does not demonstrate a general psychological theory of taste, nor does it prove that a sufficiently expressive single vector could never support equivalent computations.

The narrower observation is more useful: when models need to account for heterogeneous interests, changing behaviour or candidate-dependent relevance, researchers often find value in preserving or reconstructing distinctions that a single unconditional summary does not make explicit.

Context-aware recommendation reaches a similar conclusion from another direction. Rather than modelling relevance solely as a relationship between a user and an item, context-aware approaches incorporate additional conditions (such as time, activity or situation), because the appropriateness of the same item for the same person can vary with circumstances. [9]

Research on human decision-making makes the interpretation of observed preference still more delicate. Tversky and Simonson showed that the relative attractiveness of alternatives can depend on the composition of the choice set itself. Bettman, Luce and Payne reviewed evidence that consumer preferences are often not fully specified in advance but are partly constructed during decision processes, with the strategy used depending on the task. Slovic described related evidence from preference reversals and elicitation effects. [10][11][12]

Here too, restraint matters. This literature does not imply that stable taste is an illusion, or that every preference is reconstructed from scratch whenever a decision is made. Long-lived dispositions and recurring patterns clearly coexist with situational effects. What the evidence makes difficult to defend is a simpler assumption: that every observed choice can be read as a direct, context-free measurement of a stable underlying preference.

For computational models, that distinction is consequential. An observation is evidence about preference. It is not necessarily preference itself.

Interpretation

Consider what happens when a behavioural history is compressed into a representation optimized to predict a future interaction.

A system may observe that someone repeatedly chooses understated clothing, later spends several weeks exploring highly expressive pieces, buys an expensive item in a category where their normal spending is modest, and saves products that differ substantially from their previous purchases. For a recommendation task, the most useful response may simply be to adjust the user’s representation until the next-item prediction improves.

A broader model of the person encounters additional questions.

The expressive pieces might indicate an emerging preference, but they might also belong to a particular occasion. The expensive purchase might reveal willingness to spend in that category, or it might have been a gift. Saved products can indicate genuine attraction, comparison behaviour, aspiration, research for someone else or curiosity that never becomes preference.

Those possibilities cannot always be resolved from the observation itself.

What matters is not that compression occurs (every practical model compresses in some form) but which distinctions become difficult to recover once evidence has been combined.

Two observations that point in different directions may be noise. They may also belong to different contexts. They may represent different facets of the same person’s taste, or a transition from one state to another. One may reflect preference while the other reflects a constraint. In other cases, there may simply not be enough evidence to decide.

For a ranking system, resolving this ambiguity may be unnecessary. The system can optimize expected relevance without producing a coherent account of why the evidence differs.

For a system expected to reason about the person, the ambiguity itself can be useful information.

This suggests a distinction between prediction and interpretation. Prediction asks which outcome is likely given available evidence. Interpretation asks what the evidence supports us in believing about the person, and under which conditions that belief should apply.

The two objectives overlap, but they are not identical.

A system can achieve strong predictive accuracy while remaining agnostic about why particular observations matter. In many applications that is entirely appropriate. Problems arise only when predictive representations are asked to support broader conclusions than the objective that produced them was designed to preserve.

This is particularly important when apparently contradictory observations are involved. Averaging can be an effective statistical operation, but the average of two modes does not always describe either mode well.

Someone who prefers one style in professional settings and another in private settings does not necessarily have a weak preference for a compromise between them. Someone whose taste has genuinely changed is not necessarily best represented by a midpoint between past and present. And someone purchasing for another person should not automatically have that behaviour assimilated into their own preference profile.

The examples are simple, but the underlying problem is general: evidence can differ not only in value, but in meaning.

The same issue appears when moving across tasks.

A representation optimized to retrieve a hundred plausible items may need to preserve different information from one used to explain why an unfamiliar object could be relevant, determine whether an apparent preference is persistent, or decide how much evidence from one domain should influence another.

There is no requirement that all of these computations share a single representation. In fact, modern AI systems routinely construct intermediate representations that depend on the problem currently being solved.

The more important question is what lies upstream of those projections.

If potentially meaningful distinctions have already been collapsed, a downstream model can only recover them when enough residual information happens to remain encoded. If they have been preserved in some computationally accessible form, different tasks can decide for themselves which distinctions matter.

This is one reason we think uncertainty deserves a place in preference representation. Sparse or conflicting evidence does not always justify a single resolved conclusion. A model may sometimes need to preserve competing interpretations, different levels of confidence, or simply the fact that something remains unknown.

That is not a failure of modelling. It can be a more faithful description of what the evidence supports.

TasteGraph perspective

Our interest is not in replacing vectors with a different fashionable primitive.

The distinction we care about is between a model and a projection from that model.

An AI system may construct a vector describing what is relevant about a person for a particular retrieval task. Another representation may be more useful for ranking, comparison or reasoning in a specific situation. These task-dependent representations can be extremely effective, and there is no reason to expect embeddings to become less important as models improve.

But the representation produced for one computation does not need to become the identity of the person inside the system.

The person is not the projection.

For a more persistent model of taste, we think the useful question is therefore not “What is the best user embedding?” in isolation. It is:

What distinctions about preference must remain available so that an AI system can reason appropriately across different situations and over time?

Three requirements recur.

Taste appears to contain structure: observations form patterns, but not necessarily one undifferentiated pattern.

Taste is contextual: the meaning or relevance of an observation can depend on the situation in which it was produced or the situation in which it is being applied.

Taste evolves: evidence arrives at different moments, some patterns persist, and others change.

These properties do not prescribe an implementation. They constrain what a sufficiently general representation should make possible.

The difference is important.

A computational representation is always selective. That is its purpose. The design question is whether it discards information that the system may later need to distinguish.

For narrow prediction problems, substantial compression can be exactly the right choice. For an AI system expected to reason persistently about a person, across new tasks, changing circumstances and evidence accumulated over time, we think the standard should be different.

The system should be able to form compact representations when compact representations are useful without assuming that any one of them is the complete account of the person.

That is the intended meaning of the title.

Taste can be represented computationally without being reducible conceptually to a single point.