# Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization

[Meta](https://yomu.fyi/company/meta) · By Yuhui Ouyang, Di Wang, Sreedal Menon, Jie Tian · Jul 15, 2026

## Summary

Meta developed Hierarchical Interest Representation as an upstream representation layer to optimize deep funnel ad ranking across its platforms. The architecture organizes interactions among users, advertisers, and products into a heterogeneous, time-decayed graph and projects sparse raw edges into multi-level super-graphs of latent interest primitives. To encode these structures, a transformer-based hierarchical encoder incorporates multimodal product semantics processed via language models alongside topology-aware attention biases. Computation scales across billions of entities by pairing deep hash ID embeddings with FlexAttention to evaluate graph biases on the fly without materializing full matrices. Self-supervised cross-view distillation trains student representations against broader teacher graph views to yield universal embeddings and interest tokens.

## Takeaways

- Hierarchical Interest Representation compresses sparse, dynamic user-ad graphs into multi-tier super-graphs of stable latent interest primitives across varied granularities.
- The system keeps ID memory bounded across tens of billions of entities using deep hash embeddings and utilizes FlexAttention to compute pairwise graph-structural biases on the fly.
- Multimodal advertiser and catalog data processed through vision and language models are fused directly into graph nodes, enabling semantic generalization to rare or unseen entities.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [LLMs](https://yomu.fyi/topic/llm), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Recommendation Systems](https://yomu.fyi/topic/recommendation-systems)

[Read original post](https://engineering.fb.com/2026/07/15/ai-research/exploring-hierarchical-interest-representation-for-meta-ads-deep-funnel-optimization)