Preserving Future Choice: A Better Way to Evaluate Recommenders
The clearest answer to the title's question is conditional but direct: optimizing recommendations can narrow future choice when each output becomes part of the evidence used to produce the next output, and when the objective rewards immediate relevance or popularity more strongly than exposure breadth. Personalization is not the sole cause. The narrowing emerges from the coupling between ranking, visibility, user interaction, and retraining. This final synthesis turns that mechanism into a practical standard for judging recommenders: ask not only whether the current list is accurate, but what kinds of options the system keeps making possible over time, for whom, and with what evidence.
Quality is a trajectory, not a single score
A recommender does not simply choose items for a stable user. It participates in a changing environment. A ranked list allocates attention. Attention creates some interactions and withholds others. Those interactions become training data, and the next model learns from a record partly shaped by the previous model. When popular items receive repeated exposure, popularity can look like independent evidence of relevance. When less-visible items receive fewer opportunities to be evaluated, their absence can be misread as lack of interest. The result is a path-dependent system: later options depend partly on earlier rankings.
That mechanism explains why the title is not an argument against relevance. Relevance is useful. A system that ignores what a person finds meaningful is not preserving choice; it is imposing noise. The problem appears when relevance is defined so narrowly that familiar response becomes the only evidence of value, or when the system treats the observed history as a complete expression of preference even though the history was produced under unequal exposure. The relevant distinction is between a system that uses the past to make the next choice more useful and one that uses the past to make the future more predictable by shrinking it.
The dynamic study makes the difference visible through user trajectories. Niche-focused users begin with stronger long-tail exposure, entropy, and coverage, yet generally lose much of that advantage over repeated simulated cycles. Their calibration deteriorates most rapidly and their accuracy often weakens proportionally more. Popular-focused users can gain or preserve ranking accuracy because their histories align with mainstream content, while remaining confined to low diversity and little long-tail exposure. Diverse-focused users tend to move toward the popular-heavy pattern. These results do not say that every user loses relevance or that every model follows an identical path. They show why a system average is inadequate: an aggregate can remain stable while the system becomes less aligned with the users whose histories are most varied.
A more adequate evaluation asks four connected questions.
Is the list relevant now? Ranking measures such as nDCG or MAP remain necessary. They indicate whether items that match the evaluation's definition of relevance appear in useful positions. Removing accuracy from the objective would make it impossible to distinguish meaningful exploration from arbitrary substitution.
Does the output remain calibrated to the user and the catalog? Measures such as ΔGAP compare the popularity of recommendations with the popularity distribution in a user's history. APRI describes the popularity of the recommended items. These measures help reveal whether the system is quietly moving users toward the head even when the ranking still looks plausible.
What breadth is actually being exposed? APLT, entropy, LTC, ILD, coverage, and novelty observe different levels of exposure. APLT is closer to an individual user's long-tail experience; entropy and coverage describe broader concentration or catalog representation; ILD concerns dissimilarity within a list.