Pin-SCALE: Cascading, alignment, and deployment of engagement-aware semantic IDs¶

Core engagement-aware residual codebooks and cascade similarity.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Pinterest.
First public release	2026-07-19
Author code	No official author code linked by the paper as checked on 2026-08-09.
Local adapter or method key	`pin-scale`
Local reproduction code	`src/auto_research/reproductions/pin_scale/`

Paper method and evidence¶

Keep residual SID hierarchy via cascading pooling, encode engagement into tokenization, and use query–item and item–item contrastive views to align IDs with engagement prediction for dense retrieval.

flowchart LR
  A[Engagement-aware tokenization] --> B[Residual codebook cascade]
  B[Residual codebook cascade] --> C[Multi-view contrastive alignment]

Original paper figure¶

Figure shows SID hierarchy and alignment. From the original paper; copyright remains with the authors.

Core equation¶

\[ c_\ell=\arg\min_j w_i\lVert r_\ell-e_{\ell,j}\rVert_2^2,\qquad \mathcal L=\mathcal L_{\mathrm{rank}}+ \lambda_q\mathcal L_{q,i}+\lambda_i\mathcal L_{i,i}. \]

Reported results¶

Deployed on Pinterest Closeup, Home Feed and Search: fresh Repin +3.67%, DAU +0.05%, billions of daily requests.

Local reproduction¶

auto-research reproduce --paper pin-scale --dataset-dir data --seed 42

Unweighted semantic/content ranker vs engagement-weighted three-level SID: NDCG@10 0.03540→0.04022 (+13.61%), fresh Hit@10 +50.00% on a very small denominator. Metrics. The SIGIR program is the verified primary source, not an arXiv record. Public popularity/transitions proxy Pinterest engagement.