Idea
A graph-based embedding platform enhancing ad ranking accuracy by integrating onsite and offsite user interaction data for advertisers and platforms
Research Paper
Core Innovation
This paper introduces TransRA, a knowledge graph embedding model that constructs a heterogeneous graph combining onsite ad interactions with opt-in offsite conversion data. It uses an attention-based finetuning method to effectively integrate these embeddings, improving ad ranking performance beyond traditional models. This approach uniquely leverages both onsite and offsite user behavior for better ad relevance.
Market Size (TAM)
$20–50B TAM, $2–10B SAM; assumption: large digital advertising market with growing demand for advanced targeting and conversion tracking.
Potential Customers & Pain Points
- Digital Advertising Platforms Needing Better Ad Ranking Accuracy
- Advertisers Seeking Higher CTR and Lower CPC
- Marketing Analytics Firms Requiring Richer User Interaction Insights
Business Model
Licensing the embedding platform as a SaaS API to ad platforms and marketing analytics providers with tiered pricing based on data volume and query usage
Competitive Landscape
- Google Ads
- Facebook Ads
- The Trade Desk
Implementation Challenges
- Data Privacy and User Consent Constraints
- Integration Complexity with Existing Ad Systems
- Scalability of Large-Scale Graph Embeddings
Validation Strategy
- Deploy pilot integration with select ad platforms to measure CTR and CPC improvements
- Conduct A/B testing comparing TransRA-enhanced rankings versus baseline models
- Gather user feedback and iterate on attention-based finetuning for optimization
Research Paper Overview
Entity Representation Learning Through Onsite-Offsite Graph for Pinterest Ads
Summary
This paper presents a novel approach to improve Ads ranking models by constructing a large-scale heterogeneous graph combining users' onsite ad interactions and opt-in offsite conversion activities. It introduces TransRA, a Knowledge Graph Embedding model, and an attention-based finetuning method to integrate these embeddings effectively, resulting in significant improvements in CTR and CPC metrics deployed at Pinterest.