TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising

2026-07-28

AI Summary

This work addresses the challenge of long-cycle conversion rate prediction in online advertising, where conversions are observed asynchronously with clicks. The authors propose TWICE, a novel framework that introduces a dual-clock, dual-window modeling paradigm. It decomposes the conversion rate into the probability of conversion within a target window and the grouped delay cumulative distribution function (CDF), leveraging complementary supervisory signals from both click and conversion clocks for joint learning. By fixing the click timestamp and predicting the conversion rate for queued impressions, TWICE effectively models heterogeneous delay distributions, supports arbitrary prediction horizons, and eliminates the need for historical backfilling. Evaluated on both public and industrial datasets, TWICE significantly outperforms existing baselines. In A/B tests on Kuaishou, it achieved gains of 2.486% in expected revenue, 1.858% in actual revenue, and 2.061% in conversion count, leading to full-scale deployment.

Abstract

Long-horizon conversion prediction under delayed feedback creates a two-clock, two-window learning problem in online advertising. A short base observation window releases recent clicks on the click clock before their outcomes mature, whereas conversions continue to arrive on the conversion clock throughout a longer target conversion window. The click clock provides timely but partially observed status supervision. The conversion clock reveals long-tail delays, but the delay composition within an arrival-time slice is weighted by historical click cohorts with different traffic volumes and target-window conversion rates. We present TWICE, a framework that factorizes long-horizon post-click conversion rate (CVR) into a target-window conversion probability and a grouped elapsed-delay cumulative distribution function (CDF). The two clocks provide complementary supervision. Click-clock records train the target-window CVR head through a current-status likelihood over the base observation window. Newly arrived conversions train the delay model on the conversion clock. To account for the cohort mixture, TWICE uses fixed click-time predicted CVR (pCVR) mass as cohort exposure in an arrival-conditioned likelihood. This accounts for differences in cohort traffic and conversion propensity. The resulting aggregate records are self-contained. A single learned CDF produces monotone predictions for all requested horizons up to the target conversion window. Serving requires neither historical lookup nor convolution. Experiments on a public benchmark and an industrial advertising dataset demonstrate the effectiveness of TWICE. In an online A/B test in Kwai's advertising system, TWICE increased expected revenue, revenue, and conversions by 2.486%, 1.858%, and 2.061%, respectively. It was subsequently deployed to full traffic.

Problem

• delayed feedback
• long-horizon conversion prediction
• two-clock learning
• online advertising
• conversion rate estimation

Innovation

• delayed feedback
• conversion rate prediction
• two-clock learning
• cumulative distribution function
• online advertising