Historical Data Reuse for Precise CVR Prediction during E‑commerce Promotions

This article presents the work of Alibaba’s Advertising Ranking team on accurate conversion‑rate (CVR) estimation during large‑scale e‑commerce promotion periods. Traditional CVR models suffer significant performance drops during promotions, prompting the definition of a high‑value problem: precise CVR prediction in promotion cycles.

The authors propose the Historical Data Reuse (HDR) algorithm, which reuses fully labeled historical promotion data to fine‑tune the production CVR model. HDR was deployed for the 2022 Double‑11 promotion, delivering notable business gains (RPM +9%, CVR +16%, ROI +11%). The related paper has been accepted at KDD 2023.

Problem Analysis : Promotion periods cause abrupt user‑behavior shifts, violating the i.i.d. assumption of standard models. Delayed feedback further complicates CVR modeling because conversions may occur days after clicks.

Data Retrieval : Each day is represented by a feature vector composed of recent CVR values (including incomplete attribution) and the exposure share of representative product categories. Cosine similarity between the current day’s vector and historical days identifies the most similar past promotions for reuse.

Model Fine‑Tuning : The retrieved historical data are used to fine‑tune the current model rather than replace it. Two key challenges are addressed:

Automatic selection of similar historical days.

Correcting distribution bias between historical and target promotions.

TransBlock Module : To avoid over‑fitting (the “One‑Epoch” problem), a lightweight TransBlock layer is added on top of the base model. During fine‑tuning, TransBlock parameters are updated with a larger learning rate, while the base model parameters receive a much smaller learning rate, preserving the original model’s knowledge.

Distribution‑Correction (DSC) : An importance‑weighted empirical risk minimization framework adjusts the loss on historical data using estimated importance weights. The weights are derived from label‑shift assumptions and a simple unsupervised estimator of the target promotion’s CVR distribution.

Experiments : Offline AUC and PC​OC metrics show that HDR outperforms baseline methods (DEFER, ES‑DFM, DEFUSE, etc.) by up to 10.3% AUC improvement and achieves near‑ideal PC​OC (0.99). Ablation studies confirm the necessity of both DSC and TransBlock. Online deployment during Double‑11 demonstrated consistent RPM, CVR, and ROI improvements across product categories.

Conclusion : HDR provides a practical, data‑centric solution for the CVR degradation problem in promotion cycles, with potential applicability to other domains (search, recommendation, news, short‑video) where user behavior shifts abruptly.