Click A, Buy B: Cross-Item Conversion Attribution¶

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information¶

Field	Value
Paper	arXiv 2507.15113
Affiliations	Pinterest
First public release	2025-07-20 (arXiv v1)
Authors' code	No official author code linked as of 2026-08-09.
Adapter / method	`click-a-buy-b`
Local code	`src/auto_research/reproductions/click_a_buy_b/`

Original paper¶

Background and contribution¶

Last click assumes the same item is clicked and bought. Separate click-A-buy-A and click-A-buy-B tasks, with taxonomy-based collaborative weights across items.

flowchart LR
  C[Click A] --> A[CABA head] --> F[Conversion score]
  C --> B[CABB head] --> W[Collaborative weighting] --> F
  T[Taxonomy] --> W

Key figure from the paper¶

Original Figure 1: Click A, Buy B. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ \hat p(B|A,u)=w_{AB}^{tax}\,\hat p_{\mathrm{CABB}}(B|A,u),\qquad \mathcal L=\mathcal L_{\mathrm{CABA}}+\lambda\mathcal L_{\mathrm{CABB}}. \]

Paper results¶

The paper reports better offline cross-item attribution and +0.25% in the primary Pinterest online business metric.

Local reproduction¶

Adjacent movies proxy click and buy. Versus single-item last click: NDCG@10 0.07514→0.10045 (+33.70%), Hit@10 0.14048→0.15714, head share 0.21762→0.14119. Metrics.

auto-research reproduce --paper click-a-buy-b --dataset-dir data --seed 42

Reproduction limits¶

Watching consecutive movies is not a purchase. No real exposures, attribution window, auction, or Pinterest taxonomy.