Hybrid Contrastive Constraints for Multi-Scenario Ad Ranking

TL;DR

This paper introduces HC^2, a hybrid contrastive learning method that models cross-scenario relations in multi-scenario ad ranking, improving the ability to learn shared and scenario-specific knowledge.

Contribution

The paper proposes a novel hybrid contrastive learning framework with specialized losses and sample strategies for multi-scenario ad ranking, addressing cross-scenario relation modeling.

Findings

HC^2 outperforms baseline methods in offline evaluations.

HC^2 shows significant improvements in online A/B testing.

The approach effectively captures commonalities and differences across scenarios.

Abstract

Multi-scenario ad ranking aims at leveraging the data from multiple domains or channels for training a unified ranking model to improve the performance at each individual scenario. Although the research on this task has made important progress, it still lacks the consideration of cross-scenario relations, thus leading to limitation in learning capability and difficulty in interrelation modeling. In this paper, we propose a Hybrid Contrastive Constrained approach (HC^2) for multi-scenario ad ranking. To enhance the modeling of data interrelation, we elaborately design a hybrid contrastive learning approach to capture commonalities and differences among multiple scenarios. The core of our approach consists of two elaborated contrastive losses, namely generalized and individual contrastive loss, which aim at capturing common knowledge and scenario-specific knowledge, respectively. To adapt contrastive learning to the complex multi-scenario setting, we propose a series of important improvements. For generalized contrastive loss, we enhance contrastive learning by extending the contrastive samples (label-aware and diffusion noise enhanced contrastive samples) and reweighting the contrastive samples (reciprocal similarity weighting). For individual contrastive loss, we use the strategies of dropout-based augmentation and {cross-scenario encoding} for generating meaningful positive and negative contrastive samples, respectively. Extensive experiments on both offline evaluation and online test have demonstrated the effectiveness of the proposed HC$^2$ by comparing it with a number of competitive baselines.