Retrievable Domain-Sensitive Feature Memory for Multi-Domain Recommendation

TL;DR

This work reframes multi-domain ad ranking through the lens of feature distributions, identifying domain-sensitive features as those with large inter-domain distribution gaps and influential effects on predictions. It introduces a Domain-Sensitive Feature Attribution method to select these features and a Domain-Sensitive Feature Memory with a two-tower plus cross-attention architecture to retrieve domain-specific signals. The approach demonstrates superior offline performance, robust ablations, and favorable online results, while maintaining computational efficiency through linear attention. By explicitly targeting domain distinctions, the method improves tail-domain accuracy and enhances practical deployment in large-scale, multi-domain ad systems.

Abstract

With the increase in the business scale and number of domains in online advertising, multi-domain ad recommendation has become a mainstream solution in the industry. The core of multi-domain recommendation is effectively modeling the commonalities and distinctions among domains. Existing works are dedicated to designing model architectures for implicit multi-domain modeling while overlooking an in-depth investigation from a more fundamental perspective of feature distributions. This paper focuses on features with significant differences across various domains in both distributions and effects on model predictions. We refer to these features as domain-sensitive features, which serve as carriers of domain distinctions and are crucial for multi-domain modeling. Experiments demonstrate that existing multi-domain modeling methods may neglect domain-sensitive features, indicating insufficient learning of domain distinctions. To avoid this neglect, we propose a domain-sensitive feature attribution method to identify features that best reflect domain distinctions from the feature set. Further, we design a memory architecture that extracts domain-specific information from domain-sensitive features for the model to retrieve and integrate, thereby enhancing the awareness of domain distinctions. Extensive offline and online experiments demonstrate the superiority of our method in capturing domain distinctions and improving multi-domain recommendation performance.

Figures (7)

• Figure 1: Feature site_id and province show diverse inter-domain differences in feature distributions.
• Figure 2: Models neglect domain-sensitive features with the largest inter-domain distribution differences.
• Figure 3: The architecture of Domain-Sensitive Feature Memory consists of two key modules: the Extractor and Retriever.
• Figure 4: Results (AUC) of the ablation study on several variants of our domain-sensitive feature memory.
• Figure 5: Our method effectively improves the model's dependence on two target features: app_package and creative_type.