Hybrid Contrastive Constraints for Multi-Scenario Ad Ranking (HC²)

Multi‑scenario ad estimation modeling aims to train a unified model across diverse advertising scenarios to improve overall effectiveness. Existing methods, while successful in recommendation and advertising, often ignore cross‑scenario relationships, limiting learning capacity and making it difficult to model inter‑scene interactions.

This paper proposes a hybrid contrastive learning framework called HC². It introduces two carefully designed contrastive losses: a scene‑agnostic (generalized) contrastive loss that captures common knowledge across scenarios, and a scene‑specific (individual) contrastive loss that preserves knowledge unique to each scenario. To adapt contrastive learning to the complex multi‑scenario setting, the authors add label‑aware positive sampling, diffusion‑noise‑enhanced negative sampling, and an inverse‑similarity weighting scheme to reduce the impact of pseudo‑positive/negative pairs.

The overall architecture builds on a shared‑bottom network combined with scene‑specific subnetworks. The generalized loss operates on the shared representation, while the individual loss operates on the scene‑specific hidden layers. Both losses are jointly optimized with the primary prediction loss, with balancing hyper‑parameters controlling their influence.

Extensive offline experiments on two datasets—public AliExpress and an internal Alibaba‑ads dataset—cover CTR and CTCVR tasks. HC² consistently outperforms strong baselines, and ablation studies confirm the contribution of each component, especially the generalized contrastive loss. The method also brings consistent gains when applied to other multi‑scenario architectures such as MMoE, HMoE, STAR, and M2M.

Online A/B testing in Alibaba’s advertising system demonstrates practical impact, achieving +2.51% CVR and +3.72% GMV improvements.

In summary, HC² provides a general enhancement for multi‑scenario ad ranking by jointly learning shared and scenario‑specific representations through hybrid contrastive constraints, validated by both offline and online experiments.