Stepping up marketing for advertisers: Scalable lookalike audience

William Wu22 Sep 2023

Problem

Advertisers at Grab create custom audiences, and the platform must find more users who look like the advertiser's seed customers. The old system stored results in a Segmentation Platform (SegP). One lookalike audience took two working days, because inserting it into SegP took three times longer than generating it. Time and cost grew in proportion to the number of onboarded merchant-partners. Audiences were refreshed only weekly. The cost per segment made advertisers hesitate to run several audiences in one campaign.

Example from the article: an advertiser launches a campaign with a custom audience and has to wait two working days before it can be used.

Solution in brief

Represent every user and every audience as an embedding. The audience embedding comes from the seed users' embeddings, and a user belongs to an audience if the cosine score passes a stored threshold. Only audience embeddings and thresholds are stored, and the user embedding table is compressed with a hashing scheme so everything runs from memory.

Questions this case answers
  • How to expand a seed set of entities into a larger similar set cheaply and refresh it frequently when each new set takes days and costs too much to build
  • How to answer a per-entity membership query over many stored sets within a tight latency budget when the entity embedding table is too large for memory
  • How to build lookalike audiences for advertisers in minutes instead of days?
  • How to cut the cost of creating one lookalike audience when advertisers need many of them?

Tools and methods

  • Embedding-similarity lookalike — cosine score of user and audience embeddings against a threshold.
  • Shared user embeddings — reused across projects to share the cost.
  • Hash-based embedding compression — Bloom-filter-inspired, trained with mean square error.
  • TensorFlow Serving — hosts the online retrieval model.
Continue reading — it's free

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