SCALR: Translate cross-domain behavior into recommender training data

简体中文

Fidelity: core mechanism reproduction. Generates stochastic and deterministic top-k events from a learned distribution under a common downstream protocol.

Paper information

Field	Details
Paper	Paper or original source
Institution	Meta
First public release	2026-05-29 (arXiv v1)
Author code	No official author code in paper or arXiv entry (checked 2026-09-01).
Local adapter or method key	scalr
Local reproduction code	src/auto_research/reproductions/scalr/

Paper method and evidence

Target conversions are sparse even when the same users have rich other-surface behavior. SCALR decouples transfer from the model: estimate an item translation distribution from overlapping users, sample target-domain item \(i_k\sim\hat P(\cdot\mid u,j)\) for a source event, then add \((u,i_k,w_k)\) as ordinary synthetic training data. Sampling retains distribution/catalog diversity unlike repeated head-item top-k generation.

flowchart LR
  A[Overlapping users and source events] --> B[Estimate item translation]
  B[Estimate item translation] --> C[Sample target items]
  C[Sample target items] --> D[Train downstream model]

Original paper figure

Figure 3 shows cross-domain synthetic event translation. From the original paper; copyright remains with the authors.

Core equation

\[ \hat P(i\mid j)=\frac{\sum_{u\in U_{overlap}}\mathbf 1[(u,j)\in D_S]\mathbf 1[(u,i)\in D_T]}{\sum_{u\in U_{overlap}}\mathbf 1[(u,j)\in D_S]}. \]

Reported results

Multi-week live traffic shows statistically significant CVR gains about 0.14%–0.24%. Ablations favor stochastic sampling over top-k and show synthetic-volume saturation and the importance of distribution matching.

Local reproduction

Matched source events and per-event generation budget. On disjoint MovieLens genre domains, stochastic sampling raises generated coverage 2.5%→100% and lowers head share, but seed 42 NDCG@10 drops 10.25% relative to top-k. Three-seed metrics, SD and 95% CI. This is probability translation and coverage evidence only; genre does not reproduce cross-product surfaces, conversion labels, privacy pipeline or multi-source production service.

auto-research reproduce --paper scalr --dataset-dir data --seed 42