Making the Full Model Adaptive: Multi-level Domain Adaptation for Multi-Domain CTR Prediction
Multi-domain CTR prediction helps a single recommender model serve multiple domains with awareness of their relatedness, and existing methods usually add domain-specific layers on a shared model to consider domain characteristics. However, different domains may have distinct feature spaces and importance, and the shared model cannot effectively unify them and may neglect useful domain relations. In this paper, we propose a multi-level domain adaptation method for multi-domain CTR prediction. It introduces domain awareness to many critical steps in CTR prediction, including feature embedding, feature selection, and feature representation, to better bridge and fuse multi-domain signals. Concretely, we maintain a set of meta-embeddings for each feature field and compose them into domain-aware feature embeddings. We then select them in a domain-aware way to promote informative features for different domains. Finally, we use a domain-adaptive router to select proper submodels from multiple candidates to learn domain-specific representations. Extensive experiments on both public and proprietary datasets validate the effectiveness of our method. Its online deployment also achieves notable improvements over well-crafted predecessors.
PDF
