Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization

Yuhui Ouyang, Di Wang, Sreedal Menon, Jie Tian15 Jul 2026

Problem

Ads ranking teams optimize for deep-funnel outcomes. Millions of advertisers reach billions of people, but impressions are limited and deep-funnel feedback is scarce, so most user-advertiser pairs have almost no signal. New advertisers and products have no history at all. The cost is weaker ad relevance for the businesses that depend on deep-funnel results.

Example from the article: Meta models users, ads, advertisers, campaigns, products and pixels as one graph spanning billions of users, entities and interactions per month.

Solution in brief

Build an upstream representation layer. It projects the raw interaction graph into a smaller graph of learned interest primitives at several granularities, so sparse user-ad edges become denser. A graph transformer is trained with self-supervised teacher-student distillation plus engagement prediction, and nodes are enriched with LLM-processed content. The output is universal embeddings and discrete Bag-of-Meaning (BoM) tokens for ranking and retrieval.

Questions this case answers
  • How to learn entity representations from a very large bipartite interaction graph when the target feedback is too sparse to supervise directly
  • How to give rare or unseen entities useful representations by combining content features with interaction data
  • How to learn user and advertiser representations when deep-funnel conversions are too sparse to train on directly?
  • How to represent rare and unseen advertisers and products in a recommender without engagement history?

Tools and methods

  • Typed, time-decayed engagement graph — one graph over all ad entities.
  • Hierarchical graph transformer — graph structure enters as attention biases.
  • Cross-view distillation — self-supervised teacher-student training that finds interest clusters.
  • Multimodal content features — text, image and video of advertisers and products.
  • Composite quantization — turns embeddings into BoM tokens.
Continue reading — it's free

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