Choose a method

Start from the inputs you can provide. The method registry validates requires, so choosing a method is mostly a data-availability question.

Decision table

You have...	Try first	Then compare with
Warm donor scores + item content	linmap	grouped_most_popular_pers, knn_score_avg
Warm donor scores + content + val-cold fold	stacking_plus	linmap, stacking
Cold-to-warm similarity only	knn_score_avg	attention_knn, debiased_knn
Donor embeddings	linmap_emb	attention_emb, magnitude_scaling, dropoutnet
Only interactions/content, no donor scores	grouped_most_popular_pers	most_popular

Recommended benchmark set

methods:
  - name: grouped_most_popular_pers
  - name: linmap
  - name: stacking_plus
  - name: knn_score_avg

Add embedding methods only for donors that expose embeddings. CatBoost and EASE do not provide the same latent-factor interface as ALS/BPR/Two-Tower, so [EMB] methods can be skipped for them.

Interpret the result

• If linmap beats Grouped MP, the donor score structure transfers through content.
• If stacking_plus wins, validation-cold data adds useful popularity/affinity features.
• If KNN methods win, inspect whether the gain is real personalization or inherited popularity.
• If no transfer method beats Grouped MP, the cold content may not explain the donor signal.