Instacart's Carrot Ads team describes how they solved the cold-start problem for new retail media partners using Domain Adaptive Learning. By treating the Instacart Marketplace as a source domain and a partner's e-commerce site as the target domain, they transfer pre-trained embeddings and fine-tune domain-specific layers of a wide-and-deep pCTR model. Adaptation happens at two levels: the neural network level (shared embeddings, fine-tuned layers) and the training data level (feature alignment, taxonomy matching, feature trimming for latency). The approach delivers higher CTR, clicks per user, and ad revenue even with limited partner data, and outperforms models trained solely on target-domain data. Current limitations include manual schema mapping and human-in-the-loop verification; an automated Domain Adaptation Platform is planned.
Table of contents
IntroductionDomain Adaptive LearningWhat is Domain Adaptive Learning?Get Xiyu Wang ’s stories in your inboxBenefits & ChallengesModel ArchitectureDomain Adaptation At Neural Network LevelDomain Adaptation At Training Data LevelLearnings & ConclusionsReferencesSort: