CACS: Cascade Architecture for Creative Selection in Advertising

Creative selection is a direct carrier for displaying product content and conveying marketing information. Different creative variants for the same product can lead to vastly different performance. Existing pipelines place creative selection after ranking, preventing the ranking model from perceiving the optimal creative.

This paper proposes a novel Cascade Architecture for Creative Selection (CACS) that moves the creative selection module before the ranking stage, enabling joint optimization of intra‑ad creative selection and inter‑ad ranking.

The proposed solution includes three key components: (1) a classic dual‑tower structure to reduce computation cost and share creative representations with downstream ranking models; (2) a soft‑label list‑wise ranking distillation method that transfers knowledge from a strong ranking teacher to the creative selection student, focusing on relative ordering of creatives; (3) an adaptive dropout network that randomly drops ID features to encourage learning of multimodal content features, balancing memorization and generalization.

Experiments on both offline (sCTR) and online metrics demonstrate that CACS significantly improves click‑through rate (CTR) and revenue per mille (RPM) compared to random and post‑ranking creative selection, with only ~5% additional latency.

Extensive ablation studies show the effectiveness of the dual‑tower design, the distillation loss, and the adaptive dropout mechanism across low‑frequency and high‑frequency ads.

The approach has been deployed in the Taobao search advertising platform and achieved notable gains in both offline evaluations and live traffic.