Similarity Is Not Understanding

Recommendation systems have become remarkably good at finding things that resemble things. The resemblance may be perceptual, semantic or behavioral; in many modern systems, it is not something a human would necessarily describe as resemblance at all. Items and users can be related through latent factors learned from interaction data, candidate generators can estimate compatibility over enormous collections, and sequence models can use a person’s recent behavior to predict what may be relevant next.

For the purposes of this note, we use similarity in this broader sense: not simply geometric nearest-neighbor search, but learned proximity, affinity or compatibility between representations. This is deliberately expansive, because the argument is not that recommendation systems rely on a primitive notion of similarity. They do not. The question is whether increasingly powerful estimates of affinity are sufficient to resolve what an observation means about the person who produced it.

Consider someone who spends time looking at a particular jacket. The interaction might reflect an interest in its silhouette, material, colour or designer, but the relevant property could equally be non-visual: its price, suitability for an occasion, resemblance to something already owned, or usefulness as a reference for somebody else. The person may even be examining it because of a feature they dislike. The interaction establishes that the object became salient under some set of conditions. It does not establish which interpretation should generalize.

That distinction is easy to overlook because a system does not always need to resolve it in order to make a good prediction.

Established research

From similarity to learned affinity

The history of recommender systems provides many examples of useful prediction without an explicit semantic account of the person. Item-to-item collaborative filtering, described by Linden, Smith and York in the context of Amazon, related products through patterns of collective behavior rather than requiring a hand-designed description of why two products belonged together. Matrix-factorization methods later demonstrated that user-item preference could be represented effectively in learned latent spaces, outperforming classical nearest-neighbor techniques while incorporating signals such as implicit feedback and temporal effects (Linden, Smith & York, 2003; Koren, Bell & Volinsky, 2009).

Large-scale deep recommendation systems extended this approach considerably. The YouTube architecture described by Covington, Adams and Sargin separated candidate generation from ranking, using learned representations and behavioral information first to reduce an enormous corpus to a manageable set of candidates and then to estimate which of those candidates should be presented. This is not a system that merely retrieves objects visually or semantically similar to the last object viewed. It is a sophisticated predictive architecture designed around the practical constraints of industrial recommendation.

Research has also repeatedly enriched the representation of the user. MIND, for example, was motivated by the limitation of representing a person’s diverse interests with a single vector and instead learned multiple interest representations from behavioral history. ComiRec developed a related multi-interest framework for sequential recommendation, separating several interest representations before candidate retrieval and allowing the aggregation stage to trade off accuracy and diversity (Li et al., 2019; Cen et al., 2020).

These developments matter for the argument here because they rule out a simplistic critique. Modern recommender systems are not generally based on the assumption that a person has one interest, that an item has one relevant feature, or that recommendation consists of locating the closest object in a fixed space. The field has progressively introduced richer user representations, temporal structure, context, multiple interests and more powerful ranking functions.

What these improvements do not automatically resolve is the interpretation of the evidence from which those representations are learned.

Behavioral evidence is indirect evidence

This issue is especially visible in systems trained from implicit feedback.

A rating explicitly reports something about a person’s evaluation, although even ratings have well-known limitations. A click, view, dwell event, purchase or skip is less direct. Such observations are valuable because they occur naturally and at enormous scale, but they combine information about preference with information about exposure, interface position, availability, current intention and other conditions.

Hu, Koren and Volinsky’s work on collaborative filtering for implicit-feedback datasets made this distinction explicit: implicit observations provide evidence from which preference is inferred, rather than direct preference labels, and repeated observations can instead be treated as increasing confidence in the inferred signal (Hu, Koren & Volinsky, 2008).

Subsequent work has shown how the systems producing these observations can themselves influence the evidence. Schnabel and colleagues described selection biases arising both from user self-selection and from the recommender system determining which items are exposed. In ranking, Joachims, Swaminathan and Schnabel showed how position bias affects click behavior: results shown prominently receive different levels of interaction independently of their underlying relevance, making raw clicks an imperfect training signal (Schnabel et al., 2016; Joachims, Swaminathan & Schnabel, 2017).

These results concern different technical problems, and none of them implies that behavioral data is unusable. The opposite is true: implicit behavior remains one of the most important sources of information in modern recommendation. What the literature establishes is a narrower point. An observed interaction is produced by a person, an object and the conditions under which they encountered each other. Treating the interaction as evidence requires inference about what, precisely, it provides evidence for.

Similarity depends on what counts as relevant

There is an older parallel in cognitive research. Tversky’s Features of Similarity challenged the idea that human similarity judgments could always be adequately described by conventional geometric distance. Among other phenomena, he showed that judgments could depend on the weighting of common and distinctive features, could be asymmetric, and could vary with context (Tversky, 1977).

This result should not be read as a critique of modern machine-learned embeddings; Tversky was studying human judgments decades before such systems existed. It does, however, illustrate a more general problem with the word similar. Whether two objects count as similar depends on which distinctions are relevant to the comparison.

Two jackets may be close in visual appearance but far apart in price. Two hotels may be similar in location and category while offering very different social environments. Two pieces of music may be acoustically distant yet repeatedly serve the same purpose for one listener. A learned representation can make any of these relationships computationally accessible, but the representation and its objective determine which relationships are preserved.

Once a representation is given, a similarity computation can be precise. The harder question may precede that computation: which relation is relevant to interpreting the observation in front of us?

Prediction and explanation are different objectives

A related distinction appears in work on explainable recommendation. High predictive accuracy does not require that the internal dimensions of a model correspond to concepts that are legible to a person, which has motivated research into explanations based on aspects, features, counterfactuals and other forms of interpretation.

Tan and colleagues make a particularly useful distinction in their work on counterfactual explainable recommendation. An explanation can be considered from the model’s perspective (what caused the system to recommend an item) or from the user’s perspective (what could explain the user’s affinity for that item). These are connected questions, but they need not have the same answer (Tan et al., 2021).

This distinction is important because explainability does not eliminate the underlying ambiguity. A faithful explanation of a model’s decision may reveal why its score changed without establishing why the person originally interacted with the object. Conversely, a plausible explanation of the person’s behavior may be intuitively convincing without faithfully describing either the model’s reasoning or the person’s actual motivation.

A system can therefore be predictive without being interpretable, interpretable without recovering a person’s motives, and useful without resolving every ambiguity in the behavior it observes. These are different standards.

Interpretation

The literature does not establish that a recommender system must identify a psychologically correct explanation for every interaction. Nor would that be a plausible requirement. Human decisions can have multiple causes; people are not always able to articulate those causes themselves; and some preferences are shaped or constructed during the process of making a choice.

A more useful question is computational rather than psychological: what, if anything, should generalize from a piece of evidence?

Return to the jacket. If the relevant property was its oversized silhouette, then visually different items preserving that proportion might be more informative than superficially similar jackets. If the relevant property was the designer, the useful neighborhood changes. If the person was searching for something appropriate to a formal event, the observation may be evidence about a temporary objective rather than an enduring aesthetic preference. If the jacket was intended as a gift, even a perfectly accurate description of its attributes may tell us relatively little about the shopper’s own taste.

Each interpretation implies a different notion of relevant proximity and, potentially, a different update to the model of the person. Additional evidence may make some interpretations substantially more plausible than others, but that inference is not supplied by proximity alone.

This matters even when a predictive system performs well. Correlated features can allow a model to reach the right recommendation for reasons that are not stable outside the region of data in which those correlations hold. A person who repeatedly chooses a particular designer may also repeatedly choose a particular silhouette; within the observed catalog, either signal may retrieve appropriate items. The distinction becomes more consequential when the correlation breaks, for example with an unfamiliar brand, a new product category, a different budget or an item that preserves the relevant aesthetic property while changing several superficial ones.

The claim is not that an interpretable causal explanation is always required before recommendation can work. Often it is not. Rather, predictive success and resolution of semantic ambiguity are distinct achievements. A model can learn that two patterns tend to occur together without determining which part of one pattern should transfer when the other conditions change.

More expressive user representations help, but they do not make this distinction disappear. A multi-interest model can represent several behavioral modes instead of averaging them into one. A sequential model can distinguish recent activity from older behavior. An intent-oriented model can introduce latent variables that separate different behavioral patterns. These are important modelling advances, but their latent constructs need not correspond one-to-one with consciously held human intentions. Discovering several stable clusters in a behavior sequence and knowing what each cluster means for the person are not identical problems.

The same caution applies to natural-language explanations. Language can make a representation easier to inspect, revise or communicate, but fluency should not be mistaken for epistemic certainty. A system capable of producing a persuasive account of someone’s taste may still have limited evidence for parts of that account.

For a representation intended to persist over time, this distinction becomes increasingly important. An uncertain interpretation made from one event can influence later retrieval; the resulting exposure generates new behavior; that behavior becomes further evidence for the model. What begins as an ambiguous observation can therefore become reinforced through the system’s own subsequent choices. This is another reason to distinguish what was observed from what has been inferred from it.

TasteGraph perspective

Our view is that a model of taste should preserve that distinction for as long as it remains useful.

An observation is evidence, not an identity statement about the person. Looking at an object does not necessarily mean liking it; liking an object does not imply that every property of it should generalize; repeated choices do not, without further information, reveal which attributes, situations or intentions account for the regularity. The model may eventually accumulate enough evidence to make one interpretation substantially stronger than competing ones, but the strength of that interpretation should come from the evidence rather than from the need to collapse ambiguity immediately.

This does not require recovering a hidden psychological truth behind every click. We use understanding here in a narrower computational sense: the ability to maintain and update distinctions about what an observation may mean, what should generalize from it, under which conditions, and with what degree of confidence.

Under that definition, uncertainty is not a defect in the representation. Sometimes the appropriate representation of an observation is precisely that several explanations remain plausible. Later observations may distinguish them; some may remain unresolved indefinitely. A model that knows which parts of its interpretation are weakly supported can, in principle, behave differently from one that has already converted every trace into a fixed preference claim.

Similarity remains essential within such a system. Once there is sufficient evidence that a visual property, functional requirement, semantic concept or broader pattern matters, learned representations and retrieval methods provide remarkably efficient ways to search for other objects related along that dimension. Collaborative signals can reveal relationships no manually designed taxonomy would anticipate, while multimodal representations can expose useful structures across text, image and other modalities.

The distinction is therefore not between similarity and understanding as competing technologies. Similarity is one of the mechanisms through which an interpretation becomes operational. The question is what determines the relation along which similarity should matter.

This also explains why simply adding more representations does not completely solve the problem. Representing a person with several vectors is richer than representing them with one; representing a sequence is richer than discarding order; incorporating more contextual variables is richer than ignoring them. But a more expressive representation can still make an unjustified inference about what an observation means. Capacity and interpretation are related, but they are not interchangeable.

For short-lived recommendation tasks, the practical cost of this distinction may sometimes be small. If the objective is to rank a feed for the next few minutes, exploiting correlations in recent behavior can be entirely appropriate. A persistent model of a person faces a different problem. It accumulates observations across time and situations, and decisions about the meaning of those observations affect future reasoning. Under those conditions, preserving provenance, competing interpretations and uncertainty becomes more consequential.

Similarity answers a powerful computational question: given a representation, what is related to what?

A model of a person must also decide what the observed relationship is evidence for.

Understanding requires a model of why evidence matters, not only what it resembles.