Concepts

warm-transfer is easiest to understand as a post-hoc layer over an already trained recommender.

flowchart LR
    A[Warm interactions] --> B[Donor model]
    B --> C[Warm-item scores]
    D[Item content] --> E[Transfer method]
    C --> E
    E --> F[Cold-item scores]
    F --> G[Ranking metrics]

Core idea

The donor model learns personalization on warm items. A transfer method learns how item content maps to those donor scores and applies that mapping to cold-start items. The donor is not retrained.

Read next

• Why warm-transfer: the use case and mental model.
• Popularity bias & Grouped MP: the main failure mode of naive transfer.
• Method families: which methods exist and when to try them.
• Evaluation protocol: how the benchmark avoids cold-start leakage.