Empowering Carrot Ads with Domain Adaptive Learning

This post details Instacart's implementation of Domain Adaptive Learning for its Carrot Ads platform. It addresses the 'cold start' problem for new retail partners by transferring knowledge from the Instacart Marketplace. The approach involves adapting a wide and deep pCTR model architecture at both the Neural Network level (reusing and fine-tuning shared embedding layers, feature transfer, and fine-tuning specific layers) and the Training Data level (matching features between source and target domains, and feature trimming for latency). The post highlights benefits like improved performance in low-data scenarios, efficiency, and generalization, while also noting the current need for human-in-the-loop verification and the future development of an automated Domain Adaptation Platform.