Nine recommenders benchmarked on the same split, five behavioural reader segments, and a model that predicts how much someone will read next quarter.
One split, one protocol, every model. RMSE on held-out ratings; HR and NDCG against 99 popularity-matched negatives.
The trade nobody escapes: the models that rank best recommend from a narrow slice of the catalogue.
Share of all ratings held by the most-read share of titles.
Readers rate generously — the ceiling is real and every RMSE below lives with it.
The implicit signal agrees with the explicit one.
k-means on 17 behavioural features, projected to two principal components.
Named from each cluster's own centroid, not from the held-out label.
Permutation importance on the held-out users.
Readers with fewer than 10 ratings in training. NDCG@10, same protocol.
Top five from the hybrid, for six held-out readers.