Empowering Carrot Ads with Domain Adaptive Learning

Xiyu Wang, Trey Zhong4 May 2026

Problem

Instacart's Carrot Ads lets retailers run ad businesses on their own sites and apps. The ranking system must predict click-through rate (CTR) for every ad. A new site has almost no interaction history, and behavior differs between sites: grocery browsing is not like pet supplies or electronics. Training from scratch is data hungry, and the unchanged Instacart Marketplace model misses the partner's inventory and audience.

Example from the article: a retailer launches ads on its own e-commerce site that has little or no click history, and the model must work from day one.

Solution in brief

Treat the Marketplace as the source domain and the partner site as the target. Warm-start the partner's wide-and-deep pCTR model from Marketplace-pre-trained embeddings, fine-tune the later layers on the little partner data, and align features between catalogs. Prune features that do not help the target, so the model stays fast enough for real-time auctions.

Questions this case answers
  • How to bootstrap a click prediction model for a new domain with little history by transferring from a data-rich domain with a different catalog and user behavior
  • How to adapt a pre-trained ranking model to many target domains under real-time latency limits and varying feature availability
  • How to get an accurate click-through rate model for a new ads partner with almost no interaction history?
  • How to reuse a model trained on a large source platform for a smaller target site with different user behavior?

Tools and methods

  • Wide-and-deep pCTR model — predicts click probability from embedded user and product features
  • Pre-trained shared embeddings — carry shopping signals from the Marketplace to the new domain
  • Fine-tuning of later layers — adapts the model to the partner with little labeled data
  • Feature alignment across catalogs — makes source and target features mean the same thing
  • Feature trimming — drops features useless for a given partner, to meet latency limits
Continue reading — it's free

Google or a one-time email code. No password.