How Recommendations Become Their Own Training Data
This chapter follows recommender systems through one repeating cycle: a user history becomes a ranked list, the list changes what the user encounters, and the resulting interaction record becomes evidence for the next ranking. The focus is the transition between those states, because that is where popularity can become self-reinforcing. By the end, the loop should be clear enough to inspect—and its offline simulation distinct enough from live behavior that neither is mistaken for the other.
From history to a ranked list
A recommender begins with a record of past interactions. That record may include ratings, views, purchases, searches, or other signals that can be associated with a user and an item. It is not a complete account of what the person wants. It is a partial history of what they encountered, chose, ignored, or were willing to signal. Nevertheless, the system uses it to estimate preference for items the user has not yet interacted with.
The first transition is from history to prediction. A collaborative-filtering system looks for patterns across users and items: people with related interaction histories may be treated as evidence about one another, and items that attract related behavior may be scored as likely matches. Other systems use item descriptions or a mixture of signals. The dynamic study is concerned with collaborative-filtering models, but the loop is broader than one algorithmic family. Whatever the model, the system turns an incomplete record into scores for candidate items.
Those scores are not yet the user's experience. They become consequential when the system ranks candidates and selects a top-N list. Ranking creates an exposure bottleneck: the catalog may contain many plausible items, but only a small set is presented prominently. The act of choosing the top of the list therefore performs two jobs at once. It predicts what might be relevant, and it allocates attention among items that compete for future evidence.
An item that is ranked but never shown has little opportunity to produce an interaction. An item that appears repeatedly has many opportunities to be viewed, rated, purchased, or skipped. The system may interpret those later signals as evidence of preference, but the signals were produced under an exposure pattern that the system itself helped create. This is the hidden coupling between ranking and learning: the model does not merely observe a stable world of preferences; it helps determine which observations become available.
The dynamic framework represents the process as a sequence of prediction, top-N generation, evaluation, simulated consumption, and matrix update. At the end of one iteration, the updated user–item record is fed back into the next iteration. The same broad state transition can be described without any equations:
The important state is not just the list. It is the updated record that survives into the next cycle. If popular items are already overrepresented, and the system repeatedly gives them prominent positions, the next record may contain still more evidence about those items. The model can then treat popularity as confirmation of relevance. Less-visible items remain less observed, which makes their future scores harder to estimate. A recommendation that looked like a local ranking decision becomes a change in the conditions of later ranking.
There are two time scales inside that apparently simple transition. At the short time scale, the recommender chooses among candidates for one person at one moment. At the longer time scale, the accumulated choices reshape the evidence available to the model and the exposure distribution of the catalog. A single popular item in one list may not matter much. Repeatedly placing the same class of items in prominent positions can matter because the system is not only making several independent decisions; it is repeatedly selecting which decisions will be observed and recorded.
This also makes absence ambiguous. If a person does not interact with an item, the system may have little basis for knowing whether they disliked it, overlooked it, or never encountered it. A visible skip and an invisible alternative are not equivalent signals. Yet an interaction matrix can make both look like a missing positive response. The more a system relies on observed interactions without accounting for exposure, the more easily it can confuse “not chosen when shown” with “not chosen because it was never available for consideration.” The supplied studies do not solve this general causal problem, but the feedback framework makes clear why it matters: what is absent from the record can be the result of an earlier ranking decision.
Popularity can therefore reinforce itself through both positive evidence and missing evidence. A visible item can collect more interactions, while a less-visible item accumulates fewer opportunities to demonstrate relevance. The resulting data imbalance makes the next ranking appear reasonable even when the imbalance was partly produced by the previous ranking. Breaking that pattern does not require assuming that every less-popular item is valuable or that every familiar item is harmful. It requires recognizing that the loop needs an explicit reason to preserve opportunities for alternatives instead of allowing exposure to follow past exposure automatically.
What persists, and what changes, at each transition
The cycle is easier to understand when we ask what the recommender carries forward. User histories persist in the rating matrix or profile representation. Item popularity persists as the frequency with which items appear in the observed data. The recommendation process is rerun as the simulated record changes, so the updated data—not an assumed stable preference—drives the next evaluation cycle. What does not persist cleanly is the distinction between a person's underlying interest and the opportunity they were given to express it.
Consider a user who receives a popular film because it was highly ranked. If they watch it, the interaction may strengthen the model's belief that the film or similar films are suitable. If they do not see a niche film, the absence of a rating may be read as no preference, even though the user was never given a comparable opportunity to respond. The loop does not need a malicious rule saying “show only popular items.” It can arrive at a narrower pattern through ordinary updates that treat observed behavior as if it were independent of exposure.
This is why repeated visibility matters even when the algorithm is only trying to predict preference. Visibility changes the sample from which the system learns. It can make familiar content easier to confirm and unfamiliar content harder to evaluate. It can also alter the user's own behavior: repeated exposure may make an item recognizable, while the lack of exposure may keep another item outside consideration. The feedback is therefore two-sided. The model shapes the user's opportunities, and the user's response reshapes the model's evidence.
The term “feedback loop” should not be used to imply that every cycle produces the same outcome or that drift is inevitable. The direction depends on the objective, the candidate pool, the distribution of past interactions, the way new feedback is generated, and the extent to which the system corrects for popularity. A diversity-aware intervention can change the scoring stage. A different sampling rule can change which items are eligible. A user interface can make alternatives more visible even when the underlying ranking is unchanged. The lifecycle is a way to locate these choices, not a claim that one mechanism explains every platform.
The experiment's simulation makes another important distinction: repeat consumption is not the same as new consumption. For a previously rated item, the framework retrieves the user's most recent rating and adds a small amount of user-specific Gaussian noise. This treats repeated engagement as broadly anchored in an existing response while allowing natural variation. The noise prevents every repeat interaction from being an identical copy of the past, and the user's rating volatility determines how much variation is plausible within the simulation.
For an item the user has never rated, the framework uses an item-based collaborative-filtering procedure. It compares the new item with items already in the user's history, forms a similarity-weighted estimate from those ratings, and adds controlled noise. The result is clipped to the valid rating scale. In conceptual terms, repeat consumption says, “the past response remains informative but is not perfectly fixed,” while new consumption says, “the response is inferred from related items but is not treated as a direct observation.”
That separation matters because a simulation can become unrealistic if it uses the ranking model's own output as the user's response. Ranking scores are designed to order candidates; they may not have the same scale or variance as actual ratings. Reusing them as synthetic feedback would let the model write its own predictions into the data and then cite those predictions as independent confirmation. The dynamic framework instead uses a scale-consistent item-based estimate with noise. This does not make the behavior real, but it prevents one obvious form of circular amplification.
The choice also illustrates a general systems principle: when an output becomes an input, the representation of the transition matters as much as the model that produced the output. A score, a click, a rating, and a completed view are not interchangeable observations. They carry different meanings and different amounts of context. If a study compresses them into one record, the resulting loop may measure the assumptions of that compression rather than the behavior of a real user.
Where the simulated loop stops
The dynamic study repeats the process for ten iterations under a controlled evaluation protocol. It holds the user-level train–test structure stable, generates recommendation lists, measures them, simulates consumption, and updates the record. This design makes it possible to compare how accuracy, popularity alignment, long-tail exposure, and other measures change over time. It is valuable precisely because a one-shot evaluation cannot show whether an apparently acceptable list changes the next round of data.
But the simulation is not a live platform. It does not observe a person deciding whether to watch something while tired, changing interests after a conversation, avoiding a popular item on principle, switching devices, or responding to the way a list is explained. It does not reproduce every candidate-generation rule, interface choice, social influence, or commercial intervention. The study itself identifies simplified behavior, offline execution, and computational cost as limitations.
The difference is not merely a matter of realism in the everyday sense. It affects what can be inferred. A simulated trajectory can show that a particular combination of ranking model, synthetic consumption rule, and update cycle produces a measurable drift under stated conditions. It cannot establish that every real user would respond the same way, that the same magnitude would appear in a production system, or that a measured reduction in long-tail exposure necessarily becomes a reduction in meaningful human choice.
Nor should the limitation be used to dismiss feedback analysis. Live platforms are difficult places to isolate causes. If a system changes its ranking, interface, catalog, and business objective at once, an observed change in user behavior can be hard to attribute. Controlled simulation offers a way to test mechanisms under repeatable conditions. The correct interpretation is conditional: under this modeled lifecycle, with these data and synthetic responses, these trajectories become possible or likely. Stronger claims require evidence from other settings.
The most important lesson of the lifecycle is therefore about hidden coupling. A recommender's training data are not simply a neutral window onto preference. They are partly the residue of earlier decisions about visibility. The system can preserve that coupling, amplify it, or deliberately counterbalance it. To decide which is happening, we have to examine not only the average quality of a list but also whose histories are being changed, which items receive future exposure, and how those changes feed back into learning.
Once the cycle is visible, the next question becomes unequal impact. A loop that narrows exposure for everyone may still narrow it fastest for users who prefer niche material, or may preserve accuracy for mainstream users while reducing the breadth of their experience. Those differences require user groups and metrics that can reveal drift instead of hiding it inside one system-wide score.