TWICE: two clocks and two windows for long-delayed conversion learning¶

简体中文

Fidelity: core mechanism. Executes two clocks, the two-window correction path, and common evaluation, not private industrial features or serving.

Paper information¶

Field	Details
Paper	arXiv 2607.25404
Organization	Kuaishou
First public release	2026-07-28 (arXiv v1)
Authors' code	The paper did not supply official or author code when checked on 2026-08-09.
Adapter / method	`twice`
Local reproduction code	`src/auto_research/reproductions/twice/`

Original paper¶

Motivation and method¶

Learn current-status labels on the click clock separately from the conversion-clock delay CDF, then use exposure-window weights to correct labels that take a long time to mature.

flowchart LR
    X["Impressions and clicks"] --> C["Click-clock status"]
    Y["Conversion delay"] --> D["Conversion-clock CDF"]
    C --> J["Two-window likelihood"]
    D --> J
    J --> V["CVR"]

Key figure from the paper¶

Figure 2: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equation¶

\[ P(Y=1\mid x,t)=p_{\mathrm{cvr}}(x)\,F_{\mathrm{delay}}(t\mid x),\quad F(t+\Delta)\ge F(t). \]

Paper results¶

An online Kwai Ads A/B test reports expected revenue +2.486%, revenue +1.858%, and conversions +2.061%, followed by full-traffic deployment.

Local reproduction¶

Local comparison: The baseline is mature-label next-item CVR and the treatment is two-clock current-status modeling. Hit@10 increases +8.00% and NDCG@10 +14.31%.

MovieLens timestamps construct click and conversion clocks. The run fits a monotonic delay CDF and weights by exposure maturity.

auto-research reproduce --paper twice --data-root data --seed 42

See the stable result.

Reproduction boundaries¶

Public interactions stand in for ad delay. There are no production aggregate records or real revenue measurements; online numbers are from the paper only.