Stepping up marketing for advertisers: Scalable lookalike audience

This post introduces a scalable lookalike audience platform that addresses issues of long SLA, high cost, and unstable data ingestion in the previous system. The new platform leverages user and audience embeddings for efficient lookalike audience creation and retrieval. Key components include a real-time, in-memory retrieval service with a passenger embedding compression algorithm (reducing memory needs by ~90%) and an embedding-based audience creation and updating pipeline that utilizes shared passenger embeddings and a low-cost neural network for audience embeddings and thresholds. The implementation involves training a user embedding model on billions of user actions, generating audience embeddings by averaging passenger embeddings, and using cosine scores for audience identification. For serving, embeddings are stored in memory, and a compression method based on hash tricks (inspired by Bloom Filters) is used to reduce the 20GB embedding size. The impact includes a 15-minute update window for new campaigns, doubled impressions and clicks, improved CTR, conversion rate, and ROI, and a 98% decrease in lookalike audience generation costs.