Pinterest causal retrieval: using uplift to trigger shopping candidates¶

简体中文

Fidelity: core mechanism. Runs propensity estimation, two outcomes, doubly robust pseudo-outcomes, and a trigger threshold.

Paper information¶

Field	Details
Paper	arXiv 2607.14161
Organization	Pinterest
First public release	2026-07-14 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method	`causal-retrieval`
Local reproduction code	`src/auto_research/reproductions/causal_retrieval/`

Original paper¶

Motivation and method¶

Always triggering the shopping candidate generator wastes resources and can over-distribute to users without shopping intent. The paper learns multi-event outcomes and doubly robust uplift from randomized logs, selects a threshold via offline linear replay, and runs online in parallel with remote retrieval.

flowchart LR
  X["User and context"] --> P["Propensity"]
  X --> M0["Outcome t=0"]
  X --> M1["Outcome t=1"]
  P --> D["DR pseudo-outcome"]
  M0 --> D
  M1 --> D
  D --> U["Uplift policy"]
  U --> G{"Trigger shopping candidate generator?"}

Key figure from the paper¶

Figure 1: Training flow and key optimization steps. From the original paper; copyright remains with the authors.

Core equation¶

\[ \tilde\tau(x)=\hat\mu_1-\hat\mu_0+ \frac{t(y-\hat\mu_1)}{\hat e(x)} -\frac{(1-t)(y-\hat\mu_0)}{1-\hat e(x)}. \]

Paper results¶

Shopping triggers decline by up to 85% with neutral shopping sessions, total sessions +0.26%, and Pin Saves +1.10%.

Local reproduction¶

Local comparison: The baseline is a fixed noncausal transition/content/popularity retrieval ensemble. DR uplift triggers shopping on about 30% of cases. NDCG@10 improves 0.03540→0.06399, or +80.77% relative.

This large gain comes from synthetic randomized treatment on public data. It validates the causal training path, not a reproduction of Pinterest. See the metrics.

auto-research reproduce --paper causal-retrieval --dataset-dir data --seed 42

Reproduction boundaries¶

No Pinterest randomized logs, shopping sessions, or RPC costs are available. The local pseudo-treatment is used only for an identifiable mechanism experiment.