Finding Users Who Act Alike: Transfer Learning for Expanding Advertiser Audiences

Audience Look-alike Targeting is an online advertising technique in which an advertiser specifies a set of seed customers and tasks the advertising platform with finding an expanded audience of similar users. We will describe a two-stage embedding-based audience expansion model that is deployed in production at Pinterest. For the first stage we trained a global user embedding model on sitewide user activity logs. In the second stage, we use transfer learning and statistical techniques to create lightweight seed list representations in the embedding space for each advertiser. We create a (user, seed list) affinity scoring function that makes use of these lightweight advertiser representations. We describe the end-to-end system that computes and serves this model at scale. Finally, we propose an ensemble approach that combines single-advertiser classifiers with the embedding-based technique. We show offline evaluation and online experiments to prove that the expanded audience generated by the ensemble model has the best results for all seed list sizes.