Meta Engineering presents Hierarchical Interest Representation (HIR), a research system for improving deep funnel ad optimization. HIR builds an upstream representation layer over a massive heterogeneous graph of users, advertisers, products, and services. Key technical innovations include a transformer-based hierarchical encoder with bias-aware attention using FlexAttention for memory efficiency, self-supervised cross-view distillation via a teacher-student scheme, and multimodal knowledge enrichment through LLMs. The system outputs universal embeddings and Bag-of-Meaning (BoM) tokens that discretize interests for retrieval and ranking. Trained on billions of interactions, HIR is designed to plug into Meta's existing ads stack components like GEM, Andromeda, and the Adaptive Ranking Model to improve relevance for sparse deep-funnel signals.
