Overview
Recommender systems solve search overload by ranking likely relevant items, but repeated exposure also shapes the interactions used for future learning. This series traces that feedback loop, compares user-group trajectories with complementary accuracy, calibration, popularity, and diversity measures, and examines multidimensional scoring as a corrective. The evidence supports broader measured exposure in a controlled film benchmark, while a political-score inconsistency and the lack of live human-outcome evidence limit stronger claims.
Learning outcomes
- Repeated recommendation can make exposure part of the training data, so locally relevant lists may still narrow later opportunities for discovery.
- The simulated loop affects profiles unevenly: Niche-focused users generally lose calibration and long-tail breadth, while Popular-focused users stay accurate but narrow and Diverse-focused users drift toward the head.
- Accuracy measures such as MAP or nDCG need calibration, popularity, exposure, coverage, and diversity measures across groups and iterations.
- A multidimensional score improved measured ILD, coverage, and novelty while keeping MAP near baselines in one film benchmark, but its representations, candidate pool, weights, and political direction require audit.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
How can recommender systems narrow future choice?
They can narrow future choice when repeated rankings shape exposure, the resulting interactions become training data, and the objective rewards familiar or popular items more than breadth. This is a feedback mechanism, not proof that every recommender or every user follows the same trajectory.
Why can a recommender remain accurate while reducing choice?
A recommender can remain accurate while reducing choice because relevance scores can improve within a narrower, popularity-heavy region of the catalog. Measures such as calibration, long-tail exposure, entropy, and coverage reveal changes that MAP or nDCG alone may miss.
Can diversity-aware scoring prevent recommender feedback loops?
It may counter recommender feedback loops by letting breadth affect which candidates compete for exposure, but the supplied evidence does not show that it prevents live-platform narrowing. Candidate pools, representations, weights, formula direction, and repeated user response still determine the outcome.
What should be measured to preserve future choice?
Measure current relevance alongside calibration, popularity concentration, individual and system-level exposure, coverage, diversity, and group trajectories across repeated interactions. These proxy metrics can show narrowing, but they do not by themselves establish autonomy, welfare, or social tolerance.