A library recommendation model retrains nightly on click feedback that patrons can game with coordinated clicks. What training-time risk is this?

A library recommendation model that retrains nightly on click feedback patrons can game with coordinated clicks faces feedback-loop poisoning that steers civic recommender retraining. Interaction signals become training labels; adversaries who can manufacture clicks, holds, or ratings gradually shift recommendations toward chosen titles or away from others. Municipal libraries should treat continuous feedback-driven retrain as a training-time risk surface requiring rate limits, anomaly detection, reputation weighting, and delayed or gated retraining—not assuming every click is organic interest.

Overfitting to sparse nightly checkout histories during retraining is a data-sparsity generalization problem, not adversarial feedback-loop poisoning from coordinated clicks. Data leakage from future hold queues into training splits is a temporal validation problem, not an adversarial feedback-loop poisoning risk. Membership inference exposing individual patron borrowing histories is a privacy-inference risk, not a training-time poisoning path that steers recommendations. These options fail because they describe different failure modes rather than the adversarial feedback-to-training attack path.

Exam caveat: popularity bias and honest sudden demand spikes can resemble gaming; defenses should use statistical anomaly detection plus account/device clustering, not blunt suppression of all trends. Feedback-loop poisoning is distinct from a one-shot static dataset poison, though both corrupt training. Also consider privacy: logging anti-abuse signals must still respect patron privacy policy.

Operational check: monitor click velocity by title/account/IP clusters before each retrain; quarantine anomalous feedback from the training window; require human approval for large ranking shifts. A/B hold out a clean preference sample to detect steered drift. If coordinated bursts alone can crown a title overnight in production recommendations, the civic recommender's feedback loop remains poisonable.