Method families

The registry contains 17 methods. They differ mostly by which signal they transfer.

Score-space mapping

linmap learns content -> donor score vector directly. It is model-agnostic and usually the first strong transfer method to try. scale_shift starts from neighbor scores and adjusts scale/shift.

Supervised meta-methods

stacking, stacking_plus and logreg_calib use a validation-cold fold. They can combine transfer signals with popularity and affinity features, but they need val data.

Content neighbors

knn_score_avg, attention_knn and debiased_knn use cold-to-warm similarity. They are useful diagnostics, but naive score averaging often inherits popularity.

Embedding methods

linmap_emb, magnitude_scaling, embedding_avg, attention_emb and dropoutnet need donor embeddings. They work only when the donor exposes latent factors or item/user embeddings.

Baselines

random, most_popular, grouped_most_popular and grouped_most_popular_pers define the floor and the main comparison target. A transfer method that cannot beat personalized Grouped MP is probably not useful for the studied cold-start setting.